Add m-esr52 at 52.6.0

121 files changed, 48631 insertions, 0 deletions

diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon.c
new file mode 100644
index 000000000..dd569d348
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon.c
@@ -0,0 +1,390 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stddef.h>
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+
+void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h);
+void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h);
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(
+        int16x4_t dsrc0,
+        int16x4_t dsrc1,
+        int16x4_t dsrc2,
+        int16x4_t dsrc3,
+        int16x4_t dsrc4,
+        int16x4_t dsrc5,
+        int16x4_t dsrc6,
+        int16x4_t dsrc7,
+        int16x8_t q0s16) {
+    int32x4_t qdst;
+    int16x4_t d0s16, d1s16;
+
+    d0s16 = vget_low_s16(q0s16);
+    d1s16 = vget_high_s16(q0s16);
+
+    qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+    qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+    qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+    qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+    qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+    qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+    qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+    qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+    return qdst;
+}
+
+void vp9_convolve8_avg_horiz_neon(
+        uint8_t *src,
+        ptrdiff_t src_stride,
+        uint8_t *dst,
+        ptrdiff_t dst_stride,
+        const int16_t *filter_x,
+        int x_step_q4,
+        const int16_t *filter_y,  // unused
+        int y_step_q4,            // unused
+        int w,
+        int h) {
+    int width;
+    uint8_t *s, *d;
+    uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+    uint32x2_t d2u32, d3u32, d6u32, d7u32, d28u32, d29u32, d30u32, d31u32;
+    uint8x16_t q1u8, q3u8, q12u8, q13u8, q14u8, q15u8;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16;
+    uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+    int16x8_t q0s16;
+    uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+    int32x4_t q1s32, q2s32, q14s32, q15s32;
+    uint16x8x2_t q0x2u16;
+    uint8x8x2_t d0x2u8, d1x2u8;
+    uint32x2x2_t d0x2u32;
+    uint16x4x2_t d0x2u16, d1x2u16;
+    uint32x4x2_t q0x2u32;
+
+    if (x_step_q4 != 16) {
+        vp9_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride,
+                                  filter_x, x_step_q4,
+                                  filter_y, y_step_q4, w, h);
+        return;
+    }
+
+    q0s16 = vld1q_s16(filter_x);
+
+    src -= 3;  // adjust for taps
+    for (; h > 0; h -= 4) {  // loop_horiz_v
+        s = src;
+        d24u8 = vld1_u8(s);
+        s += src_stride;
+        d25u8 = vld1_u8(s);
+        s += src_stride;
+        d26u8 = vld1_u8(s);
+        s += src_stride;
+        d27u8 = vld1_u8(s);
+
+        q12u8 = vcombine_u8(d24u8, d25u8);
+        q13u8 = vcombine_u8(d26u8, d27u8);
+
+        q0x2u16 = vtrnq_u16(vreinterpretq_u16_u8(q12u8),
+                            vreinterpretq_u16_u8(q13u8));
+        d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+        d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+        d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+        d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+        d0x2u8 = vtrn_u8(d24u8, d25u8);
+        d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+        __builtin_prefetch(src + src_stride * 4);
+        __builtin_prefetch(src + src_stride * 5);
+
+        q8u16 = vmovl_u8(d0x2u8.val[0]);
+        q9u16 = vmovl_u8(d0x2u8.val[1]);
+        q10u16 = vmovl_u8(d1x2u8.val[0]);
+        q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+        src += 7;
+        d16u16 = vget_low_u16(q8u16);
+        d17u16 = vget_high_u16(q8u16);
+        d18u16 = vget_low_u16(q9u16);
+        d19u16 = vget_high_u16(q9u16);
+        q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+        q9u16 = vcombine_u16(d17u16, d19u16);
+
+        d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+        for (width = w;
+             width > 0;
+             width -= 4, src += 4, dst += 4) {  // loop_horiz
+            s = src;
+            d28u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d29u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d31u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+            __builtin_prefetch(src + 64);
+
+            d0x2u16 = vtrn_u16(vreinterpret_u16_u32(d28u32),
+                               vreinterpret_u16_u32(d31u32));
+            d1x2u16 = vtrn_u16(vreinterpret_u16_u32(d29u32),
+                               vreinterpret_u16_u32(d30u32));
+            d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                             vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+            d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                             vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+            __builtin_prefetch(src + 64 + src_stride);
+
+            q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+            q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+            q0x2u32 = vtrnq_u32(vreinterpretq_u32_u8(q14u8),
+                                vreinterpretq_u32_u8(q15u8));
+
+            d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+            d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+            q12u16 = vmovl_u8(d28u8);
+            q13u16 = vmovl_u8(d29u8);
+
+            __builtin_prefetch(src + 64 + src_stride * 2);
+
+            d = dst;
+            d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+            d += dst_stride;
+            d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+            d += dst_stride;
+            d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+            d += dst_stride;
+            d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+
+            d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+            d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+            d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+            d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+            d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+            d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+            d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+            d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+            d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+            q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16,
+                                    d18s16, d19s16, d23s16, d24s16, q0s16);
+            q2s32  = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16,
+                                    d19s16, d23s16, d24s16, d26s16, q0s16);
+            q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16,
+                                    d23s16, d24s16, d26s16, d27s16, q0s16);
+            q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16,
+                                    d24s16, d26s16, d27s16, d25s16, q0s16);
+
+            __builtin_prefetch(src + 64 + src_stride * 3);
+
+            d2u16 = vqrshrun_n_s32(q1s32, 7);
+            d3u16 = vqrshrun_n_s32(q2s32, 7);
+            d4u16 = vqrshrun_n_s32(q14s32, 7);
+            d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+            q1u16 = vcombine_u16(d2u16, d3u16);
+            q2u16 = vcombine_u16(d4u16, d5u16);
+
+            d2u8 = vqmovn_u16(q1u16);
+            d3u8 = vqmovn_u16(q2u16);
+
+            d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8),
+                               vreinterpret_u16_u8(d3u8));
+            d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                               vreinterpret_u32_u16(d0x2u16.val[1]));
+            d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                             vreinterpret_u8_u32(d0x2u32.val[1]));
+
+            q1u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+            q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+            q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+            d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+            d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+            d = dst;
+            vst1_lane_u32((uint32_t *)d, d2u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d2u32, 1);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+            q8u16 = q9u16;
+            d20s16 = d23s16;
+            q11u16 = q12u16;
+            q9u16 = q13u16;
+            d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        }
+        src += src_stride * 4 - w - 7;
+        dst += dst_stride * 4 - w;
+    }
+    return;
+}
+
+void vp9_convolve8_avg_vert_neon(
+        uint8_t *src,
+        ptrdiff_t src_stride,
+        uint8_t *dst,
+        ptrdiff_t dst_stride,
+        const int16_t *filter_x,  // unused
+        int x_step_q4,            // unused
+        const int16_t *filter_y,
+        int y_step_q4,
+        int w,
+        int h) {
+    int height;
+    uint8_t *s, *d;
+    uint8x8_t d2u8, d3u8;
+    uint32x2_t d2u32, d3u32, d6u32, d7u32;
+    uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+    uint8x16_t q1u8, q3u8;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16;
+    uint16x4_t d2u16, d3u16, d4u16, d5u16;
+    int16x8_t q0s16;
+    uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+    int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+    if (y_step_q4 != 16) {
+        vp9_convolve8_avg_vert_c(src, src_stride, dst, dst_stride,
+                                 filter_x, x_step_q4,
+                                 filter_y, y_step_q4, w, h);
+        return;
+    }
+
+    src -= src_stride * 3;
+    q0s16 = vld1q_s16(filter_y);
+    for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+        s = src;
+        d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+        s += src_stride;
+        d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+        s += src_stride;
+        d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+        s += src_stride;
+        d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+        s += src_stride;
+        d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+        s += src_stride;
+        d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+        s += src_stride;
+        d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+        s += src_stride;
+
+        q8u16  = vmovl_u8(vreinterpret_u8_u32(d16u32));
+        q9u16  = vmovl_u8(vreinterpret_u8_u32(d18u32));
+        q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+        q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+        d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+        d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d = dst;
+        for (height = h; height > 0; height -= 4) {  // loop_vert
+            d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+            s += src_stride;
+            d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+            s += src_stride;
+            d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+            s += src_stride;
+            d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+            s += src_stride;
+
+            q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+            q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+            d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+            d += dst_stride;
+            d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+            d += dst_stride;
+            d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+            d += dst_stride;
+            d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+            d -= dst_stride * 3;
+
+            d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+            d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+            d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+            d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+            d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+            d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+            d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+            d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+            __builtin_prefetch(s);
+            __builtin_prefetch(s + src_stride);
+            q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16,
+                                    d20s16, d21s16, d22s16, d24s16, q0s16);
+            __builtin_prefetch(s + src_stride * 2);
+            __builtin_prefetch(s + src_stride * 3);
+            q2s32  = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16,
+                                    d21s16, d22s16, d24s16, d26s16, q0s16);
+            __builtin_prefetch(d);
+            __builtin_prefetch(d + dst_stride);
+            q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16,
+                                    d22s16, d24s16, d26s16, d27s16, q0s16);
+            __builtin_prefetch(d + dst_stride * 2);
+            __builtin_prefetch(d + dst_stride * 3);
+            q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16,
+                                    d24s16, d26s16, d27s16, d25s16, q0s16);
+
+            d2u16 = vqrshrun_n_s32(q1s32, 7);
+            d3u16 = vqrshrun_n_s32(q2s32, 7);
+            d4u16 = vqrshrun_n_s32(q14s32, 7);
+            d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+            q1u16 = vcombine_u16(d2u16, d3u16);
+            q2u16 = vcombine_u16(d4u16, d5u16);
+
+            d2u8 = vqmovn_u16(q1u16);
+            d3u8 = vqmovn_u16(q2u16);
+
+            q1u8 = vcombine_u8(d2u8, d3u8);
+            q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+            q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+            d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+            d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+            vst1_lane_u32((uint32_t *)d, d2u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d2u32, 1);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 1);
+            d += dst_stride;
+
+            q8u16 = q10u16;
+            d18s16 = d22s16;
+            d19s16 = d24s16;
+            q10u16 = q13u16;
+            d22s16 = d25s16;
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon_asm.asm
new file mode 100644
index 000000000..4d85846f0
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_avg_neon_asm.asm
@@ -0,0 +1,302 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; VP9_FILTER_WEIGHT == 128
+    ; VP9_FILTER_SHIFT == 7
+
+    EXPORT  |vp9_convolve8_avg_horiz_neon|
+    EXPORT  |vp9_convolve8_avg_vert_neon|
+    IMPORT  |vp9_convolve8_avg_horiz_c|
+    IMPORT  |vp9_convolve8_avg_vert_c|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|vp9_convolve8_avg_horiz_neon| PROC
+    ldr             r12, [sp, #4]           ; x_step_q4
+    cmp             r12, #16
+    bne             vp9_convolve8_avg_horiz_c
+
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+vp9_convolve8_avg_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+vp9_convolve8_avg_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; slightly out of order load to match the existing data
+    vld1.u32        {d6[0]}, [r2], r3
+    vld1.u32        {d7[0]}, [r2], r3
+    vld1.u32        {d6[1]}, [r2], r3
+    vld1.u32        {d7[1]}, [r2], r3
+
+    sub             r2, r2, r3, lsl #2      ; reset for store
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vp9_convolve8_avg_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt vp9_convolve8_avg_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|vp9_convolve8_avg_vert_neon| PROC
+    ldr             r12, [sp, #12]
+    cmp             r12, #16
+    bne             vp9_convolve8_avg_vert_c
+
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+vp9_convolve8_avg_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+vp9_convolve8_avg_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    vld1.u32        {d6[0]}, [r5@32], r3
+    vld1.u32        {d6[1]}, [r8@32], r3
+    vld1.u32        {d7[0]}, [r5@32], r3
+    vld1.u32        {d7[1]}, [r8@32], r3
+
+    pld             [r7]
+    pld             [r4]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    sub             r5, r5, r3, lsl #1      ; reset for store
+    sub             r8, r8, r3, lsl #1
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             vp9_convolve8_avg_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vp9_convolve8_avg_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.c
new file mode 100644
index 000000000..5c555c458
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.c
@@ -0,0 +1,357 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stddef.h>
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+
+void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h);
+void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h);
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(
+        int16x4_t dsrc0,
+        int16x4_t dsrc1,
+        int16x4_t dsrc2,
+        int16x4_t dsrc3,
+        int16x4_t dsrc4,
+        int16x4_t dsrc5,
+        int16x4_t dsrc6,
+        int16x4_t dsrc7,
+        int16x8_t q0s16) {
+    int32x4_t qdst;
+    int16x4_t d0s16, d1s16;
+
+    d0s16 = vget_low_s16(q0s16);
+    d1s16 = vget_high_s16(q0s16);
+
+    qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+    qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+    qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+    qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+    qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+    qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+    qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+    qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+    return qdst;
+}
+
+void vp9_convolve8_horiz_neon(
+        uint8_t *src,
+        ptrdiff_t src_stride,
+        uint8_t *dst,
+        ptrdiff_t dst_stride,
+        const int16_t *filter_x,
+        int x_step_q4,
+        const int16_t *filter_y,  // unused
+        int y_step_q4,            // unused
+        int w,
+        int h) {
+    int width;
+    uint8_t *s, *d, *psrc, *pdst;
+    uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+    uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
+    uint8x16_t q12u8, q13u8, q14u8, q15u8;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16;
+    uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+    int16x8_t q0s16;
+    uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+    int32x4_t q1s32, q2s32, q14s32, q15s32;
+    uint16x8x2_t q0x2u16;
+    uint8x8x2_t d0x2u8, d1x2u8;
+    uint32x2x2_t d0x2u32;
+    uint16x4x2_t d0x2u16, d1x2u16;
+    uint32x4x2_t q0x2u32;
+
+    if (x_step_q4 != 16) {
+        vp9_convolve8_horiz_c(src, src_stride, dst, dst_stride,
+                              filter_x, x_step_q4,
+                              filter_y, y_step_q4, w, h);
+        return;
+    }
+
+    q0s16 = vld1q_s16(filter_x);
+
+    src -= 3;  // adjust for taps
+    for (; h > 0; h -= 4,
+        src += src_stride * 4,
+        dst += dst_stride * 4) {  // loop_horiz_v
+        s = src;
+        d24u8 = vld1_u8(s);
+        s += src_stride;
+        d25u8 = vld1_u8(s);
+        s += src_stride;
+        d26u8 = vld1_u8(s);
+        s += src_stride;
+        d27u8 = vld1_u8(s);
+
+        q12u8 = vcombine_u8(d24u8, d25u8);
+        q13u8 = vcombine_u8(d26u8, d27u8);
+
+        q0x2u16 = vtrnq_u16(vreinterpretq_u16_u8(q12u8),
+                            vreinterpretq_u16_u8(q13u8));
+        d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+        d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+        d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+        d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+        d0x2u8 = vtrn_u8(d24u8, d25u8);
+        d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+        __builtin_prefetch(src + src_stride * 4);
+        __builtin_prefetch(src + src_stride * 5);
+        __builtin_prefetch(src + src_stride * 6);
+
+        q8u16  = vmovl_u8(d0x2u8.val[0]);
+        q9u16  = vmovl_u8(d0x2u8.val[1]);
+        q10u16 = vmovl_u8(d1x2u8.val[0]);
+        q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+        d16u16 = vget_low_u16(q8u16);
+        d17u16 = vget_high_u16(q8u16);
+        d18u16 = vget_low_u16(q9u16);
+        d19u16 = vget_high_u16(q9u16);
+        q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+        q9u16 = vcombine_u16(d17u16, d19u16);
+
+        d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+        d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+        for (width = w, psrc = src + 7, pdst = dst;
+             width > 0;
+             width -= 4, psrc += 4, pdst += 4) {  // loop_horiz
+            s = psrc;
+            d28u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d29u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d31u32 = vld1_dup_u32((const uint32_t *)s);
+            s += src_stride;
+            d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+            __builtin_prefetch(psrc + 64);
+
+            d0x2u16 = vtrn_u16(vreinterpret_u16_u32(d28u32),
+                               vreinterpret_u16_u32(d31u32));
+            d1x2u16 = vtrn_u16(vreinterpret_u16_u32(d29u32),
+                               vreinterpret_u16_u32(d30u32));
+            d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                             vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+            d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                             vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+            __builtin_prefetch(psrc + 64 + src_stride);
+
+            q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+            q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+            q0x2u32 = vtrnq_u32(vreinterpretq_u32_u8(q14u8),
+                                vreinterpretq_u32_u8(q15u8));
+
+            d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+            d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+            q12u16 = vmovl_u8(d28u8);
+            q13u16 = vmovl_u8(d29u8);
+
+            __builtin_prefetch(psrc + 64 + src_stride * 2);
+
+            d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+            d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+            d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+            d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+            d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+            d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+            d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+            d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+            d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+            q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16,
+                                    d18s16, d19s16, d23s16, d24s16, q0s16);
+            q2s32  = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16,
+                                    d19s16, d23s16, d24s16, d26s16, q0s16);
+            q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16,
+                                    d23s16, d24s16, d26s16, d27s16, q0s16);
+            q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16,
+                                    d24s16, d26s16, d27s16, d25s16, q0s16);
+
+            __builtin_prefetch(psrc + 60 + src_stride * 3);
+
+            d2u16 = vqrshrun_n_s32(q1s32, 7);
+            d3u16 = vqrshrun_n_s32(q2s32, 7);
+            d4u16 = vqrshrun_n_s32(q14s32, 7);
+            d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+            q1u16 = vcombine_u16(d2u16, d3u16);
+            q2u16 = vcombine_u16(d4u16, d5u16);
+
+            d2u8 = vqmovn_u16(q1u16);
+            d3u8 = vqmovn_u16(q2u16);
+
+            d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8),
+                               vreinterpret_u16_u8(d3u8));
+            d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                               vreinterpret_u32_u16(d0x2u16.val[1]));
+            d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                             vreinterpret_u8_u32(d0x2u32.val[1]));
+
+            d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
+            d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
+
+            d = pdst;
+            vst1_lane_u32((uint32_t *)d, d2u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d2u32, 1);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+            q8u16 = q9u16;
+            d20s16 = d23s16;
+            q11u16 = q12u16;
+            q9u16 = q13u16;
+            d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+        }
+    }
+    return;
+}
+
+void vp9_convolve8_vert_neon(
+        uint8_t *src,
+        ptrdiff_t src_stride,
+        uint8_t *dst,
+        ptrdiff_t dst_stride,
+        const int16_t *filter_x,  // unused
+        int x_step_q4,            // unused
+        const int16_t *filter_y,
+        int y_step_q4,
+        int w,
+        int h) {
+    int height;
+    uint8_t *s, *d;
+    uint32x2_t d2u32, d3u32;
+    uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16;
+    uint16x4_t d2u16, d3u16, d4u16, d5u16;
+    int16x8_t q0s16;
+    uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+    int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+    if (y_step_q4 != 16) {
+        vp9_convolve8_vert_c(src, src_stride, dst, dst_stride,
+                             filter_x, x_step_q4,
+                             filter_y, y_step_q4, w, h);
+        return;
+    }
+
+    src -= src_stride * 3;
+    q0s16 = vld1q_s16(filter_y);
+    for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+        s = src;
+        d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+        s += src_stride;
+        d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+        s += src_stride;
+        d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+        s += src_stride;
+        d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+        s += src_stride;
+        d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+        s += src_stride;
+        d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+        s += src_stride;
+        d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+        s += src_stride;
+
+        q8u16  = vmovl_u8(vreinterpret_u8_u32(d16u32));
+        q9u16  = vmovl_u8(vreinterpret_u8_u32(d18u32));
+        q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+        q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+        d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+        d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+        d = dst;
+        for (height = h; height > 0; height -= 4) {  // loop_vert
+            d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+            s += src_stride;
+            d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+            s += src_stride;
+            d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+            s += src_stride;
+            d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+            s += src_stride;
+
+            q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+            q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+            d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+            d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+            d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+            d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+            d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+            d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+            d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+            d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+            __builtin_prefetch(d);
+            __builtin_prefetch(d + dst_stride);
+            q1s32  = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16,
+                                    d20s16, d21s16, d22s16, d24s16, q0s16);
+            __builtin_prefetch(d + dst_stride * 2);
+            __builtin_prefetch(d + dst_stride * 3);
+            q2s32  = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16,
+                                    d21s16, d22s16, d24s16, d26s16, q0s16);
+            __builtin_prefetch(s);
+            __builtin_prefetch(s + src_stride);
+            q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16,
+                                    d22s16, d24s16, d26s16, d27s16, q0s16);
+            __builtin_prefetch(s + src_stride * 2);
+            __builtin_prefetch(s + src_stride * 3);
+            q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16,
+                                    d24s16, d26s16, d27s16, d25s16, q0s16);
+
+            d2u16 = vqrshrun_n_s32(q1s32, 7);
+            d3u16 = vqrshrun_n_s32(q2s32, 7);
+            d4u16 = vqrshrun_n_s32(q14s32, 7);
+            d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+            q1u16 = vcombine_u16(d2u16, d3u16);
+            q2u16 = vcombine_u16(d4u16, d5u16);
+
+            d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
+            d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
+
+            vst1_lane_u32((uint32_t *)d, d2u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d2u32, 1);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 0);
+            d += dst_stride;
+            vst1_lane_u32((uint32_t *)d, d3u32, 1);
+            d += dst_stride;
+
+            q8u16 = q10u16;
+            d18s16 = d22s16;
+            d19s16 = d24s16;
+            q10u16 = q13u16;
+            d22s16 = d25s16;
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon_asm.asm
new file mode 100644
index 000000000..184c3ad67
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon_asm.asm
@@ -0,0 +1,280 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; VP9_FILTER_WEIGHT == 128
+    ; VP9_FILTER_SHIFT == 7
+
+    EXPORT  |vp9_convolve8_horiz_neon|
+    EXPORT  |vp9_convolve8_vert_neon|
+    IMPORT  |vp9_convolve8_horiz_c|
+    IMPORT  |vp9_convolve8_vert_c|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|vp9_convolve8_horiz_neon| PROC
+    ldr             r12, [sp, #4]           ; x_step_q4
+    cmp             r12, #16
+    bne             vp9_convolve8_horiz_c
+
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+vp9_convolve8_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+vp9_convolve8_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vp9_convolve8_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt vp9_convolve8_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|vp9_convolve8_vert_neon| PROC
+    ldr             r12, [sp, #12]
+    cmp             r12, #16
+    bne             vp9_convolve8_vert_c
+
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+vp9_convolve8_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+vp9_convolve8_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r7]
+    pld             [r4]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             vp9_convolve8_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             vp9_convolve8_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon.c
new file mode 100644
index 000000000..3a3db353e
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon.c
@@ -0,0 +1,145 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stddef.h>
+#include <arm_neon.h>
+
+void vp9_convolve_avg_neon(
+        const uint8_t *src,    // r0
+        ptrdiff_t src_stride,  // r1
+        uint8_t *dst,          // r2
+        ptrdiff_t dst_stride,  // r3
+        const int16_t *filter_x,
+        int filter_x_stride,
+        const int16_t *filter_y,
+        int filter_y_stride,
+        int w,
+        int h) {
+    uint8_t *d;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    uint32x2_t d0u32, d2u32;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8, q8u8, q9u8, q10u8, q11u8;
+    (void)filter_x;  (void)filter_x_stride;
+    (void)filter_y;  (void)filter_y_stride;
+
+    d = dst;
+    if (w > 32) {  // avg64
+        for (; h > 0; h -= 1) {
+            q0u8  = vld1q_u8(src);
+            q1u8  = vld1q_u8(src + 16);
+            q2u8  = vld1q_u8(src + 32);
+            q3u8  = vld1q_u8(src + 48);
+            src += src_stride;
+            q8u8  = vld1q_u8(d);
+            q9u8  = vld1q_u8(d + 16);
+            q10u8 = vld1q_u8(d + 32);
+            q11u8 = vld1q_u8(d + 48);
+            d += dst_stride;
+
+            q0u8 = vrhaddq_u8(q0u8, q8u8);
+            q1u8 = vrhaddq_u8(q1u8, q9u8);
+            q2u8 = vrhaddq_u8(q2u8, q10u8);
+            q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+            vst1q_u8(dst, q0u8);
+            vst1q_u8(dst + 16, q1u8);
+            vst1q_u8(dst + 32, q2u8);
+            vst1q_u8(dst + 48, q3u8);
+            dst += dst_stride;
+        }
+    } else if (w == 32) {  // avg32
+        for (; h > 0; h -= 2) {
+            q0u8 = vld1q_u8(src);
+            q1u8 = vld1q_u8(src + 16);
+            src += src_stride;
+            q2u8 = vld1q_u8(src);
+            q3u8 = vld1q_u8(src + 16);
+            src += src_stride;
+            q8u8 = vld1q_u8(d);
+            q9u8 = vld1q_u8(d + 16);
+            d += dst_stride;
+            q10u8 = vld1q_u8(d);
+            q11u8 = vld1q_u8(d + 16);
+            d += dst_stride;
+
+            q0u8 = vrhaddq_u8(q0u8, q8u8);
+            q1u8 = vrhaddq_u8(q1u8, q9u8);
+            q2u8 = vrhaddq_u8(q2u8, q10u8);
+            q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+            vst1q_u8(dst, q0u8);
+            vst1q_u8(dst + 16, q1u8);
+            dst += dst_stride;
+            vst1q_u8(dst, q2u8);
+            vst1q_u8(dst + 16, q3u8);
+            dst += dst_stride;
+        }
+    } else if (w > 8) {  // avg16
+        for (; h > 0; h -= 2) {
+            q0u8 = vld1q_u8(src);
+            src += src_stride;
+            q1u8 = vld1q_u8(src);
+            src += src_stride;
+            q2u8 = vld1q_u8(d);
+            d += dst_stride;
+            q3u8 = vld1q_u8(d);
+            d += dst_stride;
+
+            q0u8 = vrhaddq_u8(q0u8, q2u8);
+            q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+            vst1q_u8(dst, q0u8);
+            dst += dst_stride;
+            vst1q_u8(dst, q1u8);
+            dst += dst_stride;
+        }
+    } else if (w == 8) {  // avg8
+        for (; h > 0; h -= 2) {
+            d0u8 = vld1_u8(src);
+            src += src_stride;
+            d1u8 = vld1_u8(src);
+            src += src_stride;
+            d2u8 = vld1_u8(d);
+            d += dst_stride;
+            d3u8 = vld1_u8(d);
+            d += dst_stride;
+
+            q0u8 = vcombine_u8(d0u8, d1u8);
+            q1u8 = vcombine_u8(d2u8, d3u8);
+            q0u8 = vrhaddq_u8(q0u8, q1u8);
+
+            vst1_u8(dst, vget_low_u8(q0u8));
+            dst += dst_stride;
+            vst1_u8(dst, vget_high_u8(q0u8));
+            dst += dst_stride;
+        }
+    } else {  // avg4
+        for (; h > 0; h -= 2) {
+            d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 0);
+            src += src_stride;
+            d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 1);
+            src += src_stride;
+            d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 0);
+            d += dst_stride;
+            d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 1);
+            d += dst_stride;
+
+            d0u8 = vrhadd_u8(vreinterpret_u8_u32(d0u32),
+                             vreinterpret_u8_u32(d2u32));
+
+            d0u32 = vreinterpret_u32_u8(d0u8);
+            vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+            dst += dst_stride;
+            vst1_lane_u32((uint32_t *)dst, d0u32, 1);
+            dst += dst_stride;
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon_asm.asm
new file mode 100644
index 000000000..7d2453021
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve_avg_neon_asm.asm
@@ -0,0 +1,116 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_convolve_avg_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vp9_convolve_avg_neon| PROC
+    push                {r4-r6, lr}
+    ldrd                r4, r5, [sp, #32]
+    mov                 r6, r2
+
+    cmp                 r4, #32
+    bgt                 avg64
+    beq                 avg32
+    cmp                 r4, #8
+    bgt                 avg16
+    beq                 avg8
+    b                   avg4
+
+avg64
+    sub                 lr, r1, #32
+    sub                 r4, r3, #32
+avg64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    pld                 [r2, r3]
+    vld1.8              {q8-q9},   [r6@128]!
+    vld1.8              {q10-q11}, [r6@128], r4
+    vrhadd.u8           q0, q0, q8
+    vrhadd.u8           q1, q1, q9
+    vrhadd.u8           q2, q2, q10
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r4
+    subs                r5, r5, #1
+    bgt                 avg64_h
+    pop                 {r4-r6, pc}
+
+avg32
+    vld1.8              {q0-q1}, [r0], r1
+    vld1.8              {q2-q3}, [r0], r1
+    vld1.8              {q8-q9},   [r6@128], r3
+    vld1.8              {q10-q11}, [r6@128], r3
+    pld                 [r0]
+    vrhadd.u8           q0, q0, q8
+    pld                 [r0, r1]
+    vrhadd.u8           q1, q1, q9
+    pld                 [r6]
+    vrhadd.u8           q2, q2, q10
+    pld                 [r6, r3]
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg32
+    pop                 {r4-r6, pc}
+
+avg16
+    vld1.8              {q0}, [r0], r1
+    vld1.8              {q1}, [r0], r1
+    vld1.8              {q2}, [r6@128], r3
+    vld1.8              {q3}, [r6@128], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q2
+    pld                 [r6]
+    pld                 [r6, r3]
+    vrhadd.u8           q1, q1, q3
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg16
+    pop                 {r4-r6, pc}
+
+avg8
+    vld1.8              {d0}, [r0], r1
+    vld1.8              {d1}, [r0], r1
+    vld1.8              {d2}, [r6@64], r3
+    vld1.8              {d3}, [r6@64], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q1
+    pld                 [r6]
+    pld                 [r6, r3]
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d1}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 avg8
+    pop                 {r4-r6, pc}
+
+avg4
+    vld1.32             {d0[0]}, [r0], r1
+    vld1.32             {d0[1]}, [r0], r1
+    vld1.32             {d2[0]}, [r6@32], r3
+    vld1.32             {d2[1]}, [r6@32], r3
+    vrhadd.u8           d0, d0, d2
+    vst1.32             {d0[0]}, [r2@32], r3
+    vst1.32             {d0[1]}, [r2@32], r3
+    subs                r5, r5, #2
+    bgt                 avg4
+    pop                 {r4-r6, pc}
+    ENDP
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c
new file mode 100644
index 000000000..2e28cb20e
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve_neon.c
@@ -0,0 +1,82 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vpx_ports/mem.h"
+
+void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int x_step_q4,
+                        const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
+   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
+   */
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+
+  // Account for the vertical phase needing 3 lines prior and 4 lines post
+  int intermediate_height = h + 7;
+
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
+    vp9_convolve8_c(src, src_stride,
+                    dst, dst_stride,
+                    filter_x, x_step_q4,
+                    filter_y, y_step_q4,
+                    w, h);
+    return;
+  }
+
+  /* Filter starting 3 lines back. The neon implementation will ignore the
+   * given height and filter a multiple of 4 lines. Since this goes in to
+   * the temp buffer which has lots of extra room and is subsequently discarded
+   * this is safe if somewhat less than ideal.
+   */
+  vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
+                           temp, 64,
+                           filter_x, x_step_q4, filter_y, y_step_q4,
+                           w, intermediate_height);
+
+  /* Step into the temp buffer 3 lines to get the actual frame data */
+  vp9_convolve8_vert_neon(temp + 64 * 3, 64,
+                          dst, dst_stride,
+                          filter_x, x_step_q4, filter_y, y_step_q4,
+                          w, h);
+}
+
+void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h) {
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+  int intermediate_height = h + 7;
+
+  if (x_step_q4 != 16 || y_step_q4 != 16) {
+    vp9_convolve8_avg_c(src, src_stride,
+                        dst, dst_stride,
+                        filter_x, x_step_q4,
+                        filter_y, y_step_q4,
+                        w, h);
+    return;
+  }
+
+  /* This implementation has the same issues as above. In addition, we only want
+   * to average the values after both passes.
+   */
+  vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride,
+                           temp, 64,
+                           filter_x, x_step_q4, filter_y, y_step_q4,
+                           w, intermediate_height);
+  vp9_convolve8_avg_vert_neon(temp + 64 * 3,
+                              64, dst, dst_stride,
+                              filter_x, x_step_q4, filter_y, y_step_q4,
+                              w, h);
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_copy_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_copy_neon.c
new file mode 100644
index 000000000..f334abe11
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_copy_neon.c
@@ -0,0 +1,92 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stddef.h>
+#include <arm_neon.h>
+
+void vp9_convolve_copy_neon(
+        const uint8_t *src,    // r0
+        ptrdiff_t src_stride,  // r1
+        uint8_t *dst,          // r2
+        ptrdiff_t dst_stride,  // r3
+        const int16_t *filter_x,
+        int filter_x_stride,
+        const int16_t *filter_y,
+        int filter_y_stride,
+        int w,
+        int h) {
+    uint8x8_t d0u8, d2u8;
+    uint8x16_t q0u8, q1u8, q2u8, q3u8;
+    (void)filter_x;  (void)filter_x_stride;
+    (void)filter_y;  (void)filter_y_stride;
+
+    if (w > 32) {  // copy64
+        for (; h > 0; h--) {
+            q0u8 = vld1q_u8(src);
+            q1u8 = vld1q_u8(src + 16);
+            q2u8 = vld1q_u8(src + 32);
+            q3u8 = vld1q_u8(src + 48);
+            src += src_stride;
+
+            vst1q_u8(dst, q0u8);
+            vst1q_u8(dst + 16, q1u8);
+            vst1q_u8(dst + 32, q2u8);
+            vst1q_u8(dst + 48, q3u8);
+            dst += dst_stride;
+        }
+    } else if (w == 32) {  // copy32
+        for (; h > 0; h -= 2) {
+            q0u8 = vld1q_u8(src);
+            q1u8 = vld1q_u8(src + 16);
+            src += src_stride;
+            q2u8 = vld1q_u8(src);
+            q3u8 = vld1q_u8(src + 16);
+            src += src_stride;
+
+            vst1q_u8(dst, q0u8);
+            vst1q_u8(dst + 16, q1u8);
+            dst += dst_stride;
+            vst1q_u8(dst, q2u8);
+            vst1q_u8(dst + 16, q3u8);
+            dst += dst_stride;
+        }
+    } else if (w > 8) {  // copy16
+        for (; h > 0; h -= 2) {
+            q0u8 = vld1q_u8(src);
+            src += src_stride;
+            q1u8 = vld1q_u8(src);
+            src += src_stride;
+
+            vst1q_u8(dst, q0u8);
+            dst += dst_stride;
+            vst1q_u8(dst, q1u8);
+            dst += dst_stride;
+        }
+    } else if (w == 8) {  // copy8
+        for (; h > 0; h -= 2) {
+            d0u8 = vld1_u8(src);
+            src += src_stride;
+            d2u8 = vld1_u8(src);
+            src += src_stride;
+
+            vst1_u8(dst, d0u8);
+            dst += dst_stride;
+            vst1_u8(dst, d2u8);
+            dst += dst_stride;
+        }
+    } else {  // copy4
+        for (; h > 0; h--) {
+            *(uint32_t *)dst = *(const uint32_t *)src;
+            src += src_stride;
+            dst += dst_stride;
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_copy_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_copy_neon_asm.asm
new file mode 100644
index 000000000..a0bd04a35
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_copy_neon_asm.asm
@@ -0,0 +1,84 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_convolve_copy_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vp9_convolve_copy_neon| PROC
+    push                {r4-r5, lr}
+    ldrd                r4, r5, [sp, #28]
+
+    cmp                 r4, #32
+    bgt                 copy64
+    beq                 copy32
+    cmp                 r4, #8
+    bgt                 copy16
+    beq                 copy8
+    b                   copy4
+
+copy64
+    sub                 lr, r1, #32
+    sub                 r3, r3, #32
+copy64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #1
+    bgt                 copy64_h
+    pop                 {r4-r5, pc}
+
+copy32
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q2-q3}, [r0], r1
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy32
+    pop                 {r4-r5, pc}
+
+copy16
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q1}, [r0], r1
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy16
+    pop                 {r4-r5, pc}
+
+copy8
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d2}, [r0], r1
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d2}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 copy8
+    pop                 {r4-r5, pc}
+
+copy4
+    ldr                 r12, [r0], r1
+    str                 r12, [r2], r3
+    subs                r5, r5, #1
+    bgt                 copy4
+    pop                 {r4-r5, pc}
+    ENDP
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.c
new file mode 100644
index 000000000..0233877dd
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon.c
@@ -0,0 +1,61 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_idct.h"
+
+void vp9_idct16x16_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d2u8, d3u8, d30u8, d31u8;
+    uint64x1_t d2u64, d3u64, d4u64, d5u64;
+    uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, j, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 6);
+
+    q0s16 = vdupq_n_s16(a1);
+    q0u16 = vreinterpretq_u16_s16(q0s16);
+
+    for (d1 = d2 = dest, i = 0; i < 4; i++) {
+        for (j = 0; j < 2; j++) {
+            d2u64 = vld1_u64((const uint64_t *)d1);
+            d3u64 = vld1_u64((const uint64_t *)(d1 + 8));
+            d1 += dest_stride;
+            d4u64 = vld1_u64((const uint64_t *)d1);
+            d5u64 = vld1_u64((const uint64_t *)(d1 + 8));
+            d1 += dest_stride;
+
+            q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+            q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+            q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+            q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+            d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+            d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+            d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+            d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+            vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+            vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d3u8));
+            d2 += dest_stride;
+            vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+            vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d31u8));
+            d2 += dest_stride;
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon_asm.asm
new file mode 100644
index 000000000..b1fd21bb6
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_1_add_neon_asm.asm
@@ -0,0 +1,198 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vp9_idct16x16_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_idct16x16_1_add_neon(int16_t *input, uint8_t *dest,
+;                                    int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct16x16_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asr              r0, r0, #6                ; >> 6
+
+    vdup.s16         q0, r0                    ; duplicate a1
+    mov              r0, #8
+    sub              r2, #8
+
+    ; load destination data row0 - row3
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row4 - row7
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row8 - row11
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row12 - row15
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |vp9_idct16x16_1_add_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.c
new file mode 100644
index 000000000..5fa3f5c01
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon.c
@@ -0,0 +1,1332 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+static int16_t cospi_2_64 = 16305;
+static int16_t cospi_4_64 = 16069;
+static int16_t cospi_6_64 = 15679;
+static int16_t cospi_8_64 = 15137;
+static int16_t cospi_10_64 = 14449;
+static int16_t cospi_12_64 = 13623;
+static int16_t cospi_14_64 = 12665;
+static int16_t cospi_16_64 = 11585;
+static int16_t cospi_18_64 = 10394;
+static int16_t cospi_20_64 = 9102;
+static int16_t cospi_22_64 = 7723;
+static int16_t cospi_24_64 = 6270;
+static int16_t cospi_26_64 = 4756;
+static int16_t cospi_28_64 = 3196;
+static int16_t cospi_30_64 = 1606;
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+void vp9_idct16x16_256_add_neon_pass1(
+        int16_t *in,
+        int16_t *out,
+        int output_stride) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    uint64x1_t d16u64, d17u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q5s32, q6s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(in);
+    q8s16  = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q9s16  = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q10s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q11s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q12s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q13s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q14s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    // stage 3
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d18s16, d1s16);
+    q6s32 = vmull_s16(d19s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlal_s16(q5s32, d30s16, d0s16);
+    q6s32 = vmlal_s16(q6s32, d31s16, d0s16);
+
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d8s16 = vqrshrn_n_s32(q2s32, 14);
+    d9s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q5s32, 14);
+    d15s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    q2s32 = vmull_s16(d26s16, d2s16);
+    q3s32 = vmull_s16(d27s16, d2s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q15s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlsl_s16(q2s32, d22s16, d3s16);
+    q3s32 = vmlsl_s16(q3s32, d23s16, d3s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q15s32 = vmlal_s16(q15s32, d23s16, d2s16);
+
+    d10s16 = vqrshrn_n_s32(q2s32, 14);
+    d11s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q15s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 4
+    d30s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d30s16);
+    q11s32 = vmull_s16(d17s16, d30s16);
+    q0s32 = vmull_s16(d24s16, d30s16);
+    q1s32 = vmull_s16(d25s16, d30s16);
+
+    d30s16 = vdup_n_s16(cospi_24_64);
+    d31s16 = vdup_n_s16(cospi_8_64);
+
+    q3s32 = vaddq_s32(q2s32, q0s32);
+    q12s32 = vaddq_s32(q11s32, q1s32);
+    q13s32 = vsubq_s32(q2s32, q0s32);
+    q1s32 = vsubq_s32(q11s32, q1s32);
+
+    d16s16 = vqrshrn_n_s32(q3s32, 14);
+    d17s16 = vqrshrn_n_s32(q12s32, 14);
+    d18s16 = vqrshrn_n_s32(q13s32, 14);
+    d19s16 = vqrshrn_n_s32(q1s32, 14);
+    q8s16 = vcombine_s16(d16s16, d17s16);
+    q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q0s32 = vmull_s16(d20s16, d31s16);
+    q1s32 = vmull_s16(d21s16, d31s16);
+    q12s32 = vmull_s16(d20s16, d30s16);
+    q13s32 = vmull_s16(d21s16, d30s16);
+
+    q0s32 = vmlal_s16(q0s32, d28s16, d30s16);
+    q1s32 = vmlal_s16(q1s32, d29s16, d30s16);
+    q12s32 = vmlsl_s16(q12s32, d28s16, d31s16);
+    q13s32 = vmlsl_s16(q13s32, d29s16, d31s16);
+
+    d22s16 = vqrshrn_n_s32(q0s32, 14);
+    d23s16 = vqrshrn_n_s32(q1s32, 14);
+    d20s16 = vqrshrn_n_s32(q12s32, 14);
+    d21s16 = vqrshrn_n_s32(q13s32, 14);
+    q10s16 = vcombine_s16(d20s16, d21s16);
+    q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q15s16 = vaddq_s16(q6s16, q7s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    // stage 5
+    q0s16 = vaddq_s16(q8s16, q11s16);
+    q1s16 = vaddq_s16(q9s16, q10s16);
+    q2s16 = vsubq_s16(q9s16, q10s16);
+    q3s16 = vsubq_s16(q8s16, q11s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q11s32 = vmull_s16(d26s16, d16s16);
+    q12s32 = vmull_s16(d27s16, d16s16);
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+
+    q6s32 = vsubq_s32(q9s32, q11s32);
+    q13s32 = vsubq_s32(q10s32, q12s32);
+    q9s32 = vaddq_s32(q9s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q12s32);
+
+    d10s16 = vqrshrn_n_s32(q6s32, 14);
+    d11s16 = vqrshrn_n_s32(q13s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q10s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 6
+    q8s16 = vaddq_s16(q0s16, q15s16);
+    q9s16 = vaddq_s16(q1s16, q6s16);
+    q10s16 = vaddq_s16(q2s16, q5s16);
+    q11s16 = vaddq_s16(q3s16, q4s16);
+    q12s16 = vsubq_s16(q3s16, q4s16);
+    q13s16 = vsubq_s16(q2s16, q5s16);
+    q14s16 = vsubq_s16(q1s16, q6s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+    d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+    d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+    d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+    d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    // store the data
+    output_stride >>= 1;  // output_stride / 2, out is int16_t
+    vst1_u64((uint64_t *)out, d16u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d17u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d20u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d21u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d22u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d23u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d24u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
+
+void vp9_idct16x16_256_add_neon_pass2(
+        int16_t *src,
+        int16_t *out,
+        int16_t *pass1Output,
+        int16_t skip_adding,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d;
+    uint8x8_t d12u8, d13u8;
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64;
+    int64x1_t d12s64, d13s64;
+    uint16x8_t q2u16, q3u16, q4u16, q5u16, q8u16;
+    uint16x8_t q9u16, q12u16, q13u16, q14u16, q15u16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(src);
+    q8s16  = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q9s16  = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q10s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q11s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q12s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q13s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q14s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    // stage 3
+    d12s16 = vdup_n_s16(cospi_30_64);
+    d13s16 = vdup_n_s16(cospi_2_64);
+
+    q2s32 = vmull_s16(d16s16, d12s16);
+    q3s32 = vmull_s16(d17s16, d12s16);
+    q1s32 = vmull_s16(d16s16, d13s16);
+    q4s32 = vmull_s16(d17s16, d13s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d13s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d13s16);
+    q1s32 = vmlal_s16(q1s32, d30s16, d12s16);
+    q4s32 = vmlal_s16(q4s32, d31s16, d12s16);
+
+    d0s16 = vqrshrn_n_s32(q2s32, 14);
+    d1s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q1s32, 14);
+    d15s16 = vqrshrn_n_s32(q4s32, 14);
+    q0s16 = vcombine_s16(d0s16, d1s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d30s16 = vdup_n_s16(cospi_14_64);
+    d31s16 = vdup_n_s16(cospi_18_64);
+
+    q2s32 = vmull_s16(d24s16, d30s16);
+    q3s32 = vmull_s16(d25s16, d30s16);
+    q4s32 = vmull_s16(d24s16, d31s16);
+    q5s32 = vmull_s16(d25s16, d31s16);
+
+    q2s32 = vmlsl_s16(q2s32, d22s16, d31s16);
+    q3s32 = vmlsl_s16(q3s32, d23s16, d31s16);
+    q4s32 = vmlal_s16(q4s32, d22s16, d30s16);
+    q5s32 = vmlal_s16(q5s32, d23s16, d30s16);
+
+    d2s16 = vqrshrn_n_s32(q2s32, 14);
+    d3s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q4s32, 14);
+    d13s16 = vqrshrn_n_s32(q5s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    d30s16 = vdup_n_s16(cospi_22_64);
+    d31s16 = vdup_n_s16(cospi_10_64);
+
+    q11s32 = vmull_s16(d20s16, d30s16);
+    q12s32 = vmull_s16(d21s16, d30s16);
+    q4s32 = vmull_s16(d20s16, d31s16);
+    q5s32 = vmull_s16(d21s16, d31s16);
+
+    q11s32 = vmlsl_s16(q11s32, d26s16, d31s16);
+    q12s32 = vmlsl_s16(q12s32, d27s16, d31s16);
+    q4s32 = vmlal_s16(q4s32, d26s16, d30s16);
+    q5s32 = vmlal_s16(q5s32, d27s16, d30s16);
+
+    d4s16 = vqrshrn_n_s32(q11s32, 14);
+    d5s16 = vqrshrn_n_s32(q12s32, 14);
+    d11s16 = vqrshrn_n_s32(q5s32, 14);
+    d10s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    d30s16 = vdup_n_s16(cospi_6_64);
+    d31s16 = vdup_n_s16(cospi_26_64);
+
+    q10s32 = vmull_s16(d28s16, d30s16);
+    q11s32 = vmull_s16(d29s16, d30s16);
+    q12s32 = vmull_s16(d28s16, d31s16);
+    q13s32 = vmull_s16(d29s16, d31s16);
+
+    q10s32 = vmlsl_s16(q10s32, d18s16, d31s16);
+    q11s32 = vmlsl_s16(q11s32, d19s16, d31s16);
+    q12s32 = vmlal_s16(q12s32, d18s16, d30s16);
+    q13s32 = vmlal_s16(q13s32, d19s16, d30s16);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q11s32, 14);
+    d8s16 = vqrshrn_n_s32(q12s32, 14);
+    d9s16 = vqrshrn_n_s32(q13s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 3
+    q9s16  = vsubq_s16(q0s16, q1s16);
+    q0s16  = vaddq_s16(q0s16, q1s16);
+    q10s16 = vsubq_s16(q3s16, q2s16);
+    q11s16 = vaddq_s16(q2s16, q3s16);
+    q12s16 = vaddq_s16(q4s16, q5s16);
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16  = vaddq_s16(q6s16, q7s16);
+
+    // stage 4
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    q2s32 = vmull_s16(d18s16, d31s16);
+    q3s32 = vmull_s16(d19s16, d31s16);
+    q4s32 = vmull_s16(d28s16, d31s16);
+    q5s32 = vmull_s16(d29s16, d31s16);
+
+    q2s32 = vmlal_s16(q2s32, d28s16, d30s16);
+    q3s32 = vmlal_s16(q3s32, d29s16, d30s16);
+    q4s32 = vmlsl_s16(q4s32, d18s16, d30s16);
+    q5s32 = vmlsl_s16(q5s32, d19s16, d30s16);
+
+    d12s16 = vqrshrn_n_s32(q2s32, 14);
+    d13s16 = vqrshrn_n_s32(q3s32, 14);
+    d2s16 = vqrshrn_n_s32(q4s32, 14);
+    d3s16 = vqrshrn_n_s32(q5s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    q3s16 = q11s16;
+    q4s16 = q12s16;
+
+    d30s16 = vdup_n_s16(-cospi_8_64);
+    q11s32 = vmull_s16(d26s16, d30s16);
+    q12s32 = vmull_s16(d27s16, d30s16);
+    q8s32 = vmull_s16(d20s16, d30s16);
+    q9s32 = vmull_s16(d21s16, d30s16);
+
+    q11s32 = vmlsl_s16(q11s32, d20s16, d31s16);
+    q12s32 = vmlsl_s16(q12s32, d21s16, d31s16);
+    q8s32 = vmlal_s16(q8s32, d26s16, d31s16);
+    q9s32 = vmlal_s16(q9s32, d27s16, d31s16);
+
+    d4s16 = vqrshrn_n_s32(q11s32, 14);
+    d5s16 = vqrshrn_n_s32(q12s32, 14);
+    d10s16 = vqrshrn_n_s32(q8s32, 14);
+    d11s16 = vqrshrn_n_s32(q9s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    // stage 5
+    q8s16  = vaddq_s16(q0s16, q3s16);
+    q9s16  = vaddq_s16(q1s16, q2s16);
+    q10s16 = vsubq_s16(q1s16, q2s16);
+    q11s16 = vsubq_s16(q0s16, q3s16);
+    q12s16 = vsubq_s16(q7s16, q4s16);
+    q13s16 = vsubq_s16(q6s16, q5s16);
+    q14s16 = vaddq_s16(q6s16, q5s16);
+    q15s16 = vaddq_s16(q7s16, q4s16);
+
+    // stage 6
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+
+    d14s16 = vdup_n_s16(cospi_16_64);
+
+    q3s32 = vmull_s16(d26s16, d14s16);
+    q4s32 = vmull_s16(d27s16, d14s16);
+    q0s32 = vmull_s16(d20s16, d14s16);
+    q1s32 = vmull_s16(d21s16, d14s16);
+
+    q5s32 = vsubq_s32(q3s32, q0s32);
+    q6s32 = vsubq_s32(q4s32, q1s32);
+    q10s32 = vaddq_s32(q3s32, q0s32);
+    q4s32 = vaddq_s32(q4s32, q1s32);
+
+    d4s16 = vqrshrn_n_s32(q5s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    d10s16 = vqrshrn_n_s32(q10s32, 14);
+    d11s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q0s32 = vmull_s16(d22s16, d14s16);
+    q1s32 = vmull_s16(d23s16, d14s16);
+    q13s32 = vmull_s16(d24s16, d14s16);
+    q6s32 = vmull_s16(d25s16, d14s16);
+
+    q10s32 = vsubq_s32(q13s32, q0s32);
+    q4s32 = vsubq_s32(q6s32, q1s32);
+    q13s32 = vaddq_s32(q13s32, q0s32);
+    q6s32 = vaddq_s32(q6s32, q1s32);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q4s32, 14);
+    d8s16 = vqrshrn_n_s32(q13s32, 14);
+    d9s16 = vqrshrn_n_s32(q6s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 7
+    if (skip_adding != 0) {
+        d = dest;
+        // load the data in pass1
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+
+        q12s16 = vaddq_s16(q0s16, q15s16);
+        q13s16 = vaddq_s16(q1s16, q14s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q14s16 = vsubq_s16(q1s16, q14s16);
+        q15s16 = vsubq_s16(q0s16, q15s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q10s16, q5s16);
+        q13s16 = vaddq_s16(q11s16, q4s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q4s16 = vsubq_s16(q11s16, q4s16);
+        q5s16 = vsubq_s16(q10s16, q5s16);
+
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q0s16, q3s16);
+        q13s16 = vaddq_s16(q1s16, q2s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q2s16 = vsubq_s16(q1s16, q2s16);
+        q3s16 = vsubq_s16(q0s16, q3s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        d13s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q12s16 = vaddq_s16(q10s16, q9s16);
+        q13s16 = vaddq_s16(q11s16, q8s16);
+        q12s16 = vrshrq_n_s16(q12s16, 6);
+        q13s16 = vrshrq_n_s16(q13s16, 6);
+        q12u16 = vaddw_u8(vreinterpretq_u16_s16(q12s16),
+                          vreinterpret_u8_s64(d12s64));
+        q13u16 = vaddw_u8(vreinterpretq_u16_s16(q13s16),
+                          vreinterpret_u8_s64(d13s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+        d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+        d += dest_stride;
+        q8s16 = vsubq_s16(q11s16, q8s16);
+        q9s16 = vsubq_s16(q10s16, q9s16);
+
+        // store the data  out 8,9,10,11,12,13,14,15
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q8s16 = vrshrq_n_s16(q8s16, 6);
+        q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q9s16 = vrshrq_n_s16(q9s16, 6);
+        q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q2s16 = vrshrq_n_s16(q2s16, 6);
+        q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q3s16 = vrshrq_n_s16(q3s16, 6);
+        q3u16 = vaddw_u8(vreinterpretq_u16_s16(q3s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q3u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q4s16 = vrshrq_n_s16(q4s16, 6);
+        q4u16 = vaddw_u8(vreinterpretq_u16_s16(q4s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q4u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q5s16 = vrshrq_n_s16(q5s16, 6);
+        q5u16 = vaddw_u8(vreinterpretq_u16_s16(q5s16),
+                         vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q5u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        dest += dest_stride;
+        q14s16 = vrshrq_n_s16(q14s16, 6);
+        q14u16 = vaddw_u8(vreinterpretq_u16_s16(q14s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q14u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+        d += dest_stride;
+
+        d12s64 = vld1_s64((int64_t *)dest);
+        q15s16 = vrshrq_n_s16(q15s16, 6);
+        q15u16 = vaddw_u8(vreinterpretq_u16_s16(q15s16),
+                          vreinterpret_u8_s64(d12s64));
+        d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q15u16));
+        vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    } else {  // skip_adding_dest
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q0s16, q15s16);
+        q13s16 = vaddq_s16(q1s16, q14s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q14s16 = vsubq_s16(q1s16, q14s16);
+        q15s16 = vsubq_s16(q0s16, q15s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q10s16, q5s16);
+        q13s16 = vaddq_s16(q11s16, q4s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q4s16 = vsubq_s16(q11s16, q4s16);
+        q5s16 = vsubq_s16(q10s16, q5s16);
+
+        q0s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q1s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q0s16, q3s16);
+        q13s16 = vaddq_s16(q1s16, q2s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q2s16 = vsubq_s16(q1s16, q2s16);
+        q3s16 = vsubq_s16(q0s16, q3s16);
+
+        q10s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q11s16 = vld1q_s16(pass1Output);
+        pass1Output += 8;
+        q12s16 = vaddq_s16(q10s16, q9s16);
+        q13s16 = vaddq_s16(q11s16, q8s16);
+        d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+        d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+        d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+        d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+        vst1_u64((uint64_t *)out, d24u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d25u64);
+        out += 12;
+        vst1_u64((uint64_t *)out, d26u64);
+        out += 4;
+        vst1_u64((uint64_t *)out, d27u64);
+        out += 12;
+        q8s16 = vsubq_s16(q11s16, q8s16);
+        q9s16 = vsubq_s16(q10s16, q9s16);
+
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q8s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q8s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q9s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q9s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q2s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q2s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q3s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q3s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q4s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q4s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q5s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q5s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q14s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q14s16)));
+        out += 12;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q15s16)));
+        out += 4;
+        vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q15s16)));
+    }
+    return;
+}
+
+void vp9_idct16x16_10_add_neon_pass1(
+        int16_t *in,
+        int16_t *out,
+        int output_stride) {
+    int16x4_t d4s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    uint64x1_t d4u64, d5u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q6s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q15s32;
+    int16x8x2_t q0x2s16;
+
+    q0x2s16 = vld2q_s16(in);
+    q8s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q9s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q10s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q11s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q12s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q13s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q14s16 = q0x2s16.val[0];
+    in += 16;
+    q0x2s16 = vld2q_s16(in);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // stage 3
+    q0s16 = vdupq_n_s16(cospi_28_64 * 2);
+    q1s16 = vdupq_n_s16(cospi_4_64 * 2);
+
+    q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+    q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+    // stage 4
+    q1s16 = vdupq_n_s16(cospi_16_64 * 2);
+    d4s16 = vdup_n_s16(cospi_16_64);
+
+    q8s16 = vqrdmulhq_s16(q8s16, q1s16);
+
+    d8s16 = vget_low_s16(q4s16);
+    d9s16 = vget_high_s16(q4s16);
+    d14s16 = vget_low_s16(q7s16);
+    d15s16 = vget_high_s16(q7s16);
+    q9s32  = vmull_s16(d14s16, d4s16);
+    q10s32 = vmull_s16(d15s16, d4s16);
+    q12s32 = vmull_s16(d9s16, d4s16);
+    q11s32 = vmull_s16(d8s16, d4s16);
+
+    q15s32 = vsubq_s32(q10s32, q12s32);
+    q6s32 = vsubq_s32(q9s32, q11s32);
+    q9s32 = vaddq_s32(q9s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q12s32);
+
+    d11s16 = vqrshrn_n_s32(q15s32, 14);
+    d10s16 = vqrshrn_n_s32(q6s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q10s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 6
+    q2s16 = vaddq_s16(q8s16, q7s16);
+    q9s16 = vaddq_s16(q8s16, q6s16);
+    q10s16 = vaddq_s16(q8s16, q5s16);
+    q11s16 = vaddq_s16(q8s16, q4s16);
+    q12s16 = vsubq_s16(q8s16, q4s16);
+    q13s16 = vsubq_s16(q8s16, q5s16);
+    q14s16 = vsubq_s16(q8s16, q6s16);
+    q15s16 = vsubq_s16(q8s16, q7s16);
+
+    d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16));
+    d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+    d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+    d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+    d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    // store the data
+    output_stride >>= 1;  // output_stride / 2, out is int16_t
+    vst1_u64((uint64_t *)out, d4u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d5u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d20u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d21u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d22u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d23u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d24u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += output_stride;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
+
+void vp9_idct16x16_10_add_neon_pass2(
+        int16_t *src,
+        int16_t *out,
+        int16_t *pass1Output,
+        int16_t skip_adding,
+        uint8_t *dest,
+        int dest_stride) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d30s16, d31s16;
+    uint64x1_t d4u64, d5u64, d6u64, d7u64, d8u64, d9u64, d10u64, d11u64;
+    uint64x1_t d16u64, d17u64, d18u64, d19u64;
+    uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32;
+    int16x8x2_t q0x2s16;
+    (void)skip_adding;
+    (void)dest;
+    (void)dest_stride;
+
+    q0x2s16 = vld2q_s16(src);
+    q8s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q9s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q10s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q11s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q12s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q13s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q14s16 = q0x2s16.val[0];
+    src += 16;
+    q0x2s16 = vld2q_s16(src);
+    q15s16 = q0x2s16.val[0];
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // stage 3
+    q6s16 = vdupq_n_s16(cospi_30_64 * 2);
+    q0s16 = vqrdmulhq_s16(q8s16, q6s16);
+    q6s16 = vdupq_n_s16(cospi_2_64 * 2);
+    q7s16 = vqrdmulhq_s16(q8s16, q6s16);
+
+    q15s16 = vdupq_n_s16(-cospi_26_64 * 2);
+    q14s16 = vdupq_n_s16(cospi_6_64 * 2);
+    q3s16 = vqrdmulhq_s16(q9s16, q15s16);
+    q4s16 = vqrdmulhq_s16(q9s16, q14s16);
+
+    // stage 4
+    d0s16 = vget_low_s16(q0s16);
+    d1s16 = vget_high_s16(q0s16);
+    d6s16 = vget_low_s16(q3s16);
+    d7s16 = vget_high_s16(q3s16);
+    d8s16 = vget_low_s16(q4s16);
+    d9s16 = vget_high_s16(q4s16);
+    d14s16 = vget_low_s16(q7s16);
+    d15s16 = vget_high_s16(q7s16);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    q12s32 = vmull_s16(d14s16, d31s16);
+    q5s32 = vmull_s16(d15s16, d31s16);
+    q2s32 = vmull_s16(d0s16, d31s16);
+    q11s32 = vmull_s16(d1s16, d31s16);
+
+    q12s32 = vmlsl_s16(q12s32, d0s16, d30s16);
+    q5s32 = vmlsl_s16(q5s32, d1s16, d30s16);
+    q2s32 = vmlal_s16(q2s32, d14s16, d30s16);
+    q11s32 = vmlal_s16(q11s32, d15s16, d30s16);
+
+    d2s16 = vqrshrn_n_s32(q12s32, 14);
+    d3s16 = vqrshrn_n_s32(q5s32, 14);
+    d12s16 = vqrshrn_n_s32(q2s32, 14);
+    d13s16 = vqrshrn_n_s32(q11s32, 14);
+    q1s16 = vcombine_s16(d2s16, d3s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    d30s16 = vdup_n_s16(-cospi_8_64);
+    q10s32 = vmull_s16(d8s16, d30s16);
+    q13s32 = vmull_s16(d9s16, d30s16);
+    q8s32 = vmull_s16(d6s16, d30s16);
+    q9s32 = vmull_s16(d7s16, d30s16);
+
+    q10s32 = vmlsl_s16(q10s32, d6s16, d31s16);
+    q13s32 = vmlsl_s16(q13s32, d7s16, d31s16);
+    q8s32 = vmlal_s16(q8s32, d8s16, d31s16);
+    q9s32 = vmlal_s16(q9s32, d9s16, d31s16);
+
+    d4s16 = vqrshrn_n_s32(q10s32, 14);
+    d5s16 = vqrshrn_n_s32(q13s32, 14);
+    d10s16 = vqrshrn_n_s32(q8s32, 14);
+    d11s16 = vqrshrn_n_s32(q9s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    // stage 5
+    q8s16  = vaddq_s16(q0s16, q3s16);
+    q9s16  = vaddq_s16(q1s16, q2s16);
+    q10s16 = vsubq_s16(q1s16, q2s16);
+    q11s16 = vsubq_s16(q0s16, q3s16);
+    q12s16 = vsubq_s16(q7s16, q4s16);
+    q13s16 = vsubq_s16(q6s16, q5s16);
+    q14s16 = vaddq_s16(q6s16, q5s16);
+    q15s16 = vaddq_s16(q7s16, q4s16);
+
+    // stage 6
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+
+    d14s16 = vdup_n_s16(cospi_16_64);
+    q3s32 = vmull_s16(d26s16, d14s16);
+    q4s32 = vmull_s16(d27s16, d14s16);
+    q0s32 = vmull_s16(d20s16, d14s16);
+    q1s32 = vmull_s16(d21s16, d14s16);
+
+    q5s32 = vsubq_s32(q3s32, q0s32);
+    q6s32 = vsubq_s32(q4s32, q1s32);
+    q0s32 = vaddq_s32(q3s32, q0s32);
+    q4s32 = vaddq_s32(q4s32, q1s32);
+
+    d4s16 = vqrshrn_n_s32(q5s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    d10s16 = vqrshrn_n_s32(q0s32, 14);
+    d11s16 = vqrshrn_n_s32(q4s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q0s32 = vmull_s16(d22s16, d14s16);
+    q1s32 = vmull_s16(d23s16, d14s16);
+    q13s32 = vmull_s16(d24s16, d14s16);
+    q6s32 = vmull_s16(d25s16, d14s16);
+
+    q10s32 = vsubq_s32(q13s32, q0s32);
+    q4s32 = vsubq_s32(q6s32, q1s32);
+    q13s32 = vaddq_s32(q13s32, q0s32);
+    q6s32 = vaddq_s32(q6s32, q1s32);
+
+    d6s16 = vqrshrn_n_s32(q10s32, 14);
+    d7s16 = vqrshrn_n_s32(q4s32, 14);
+    d8s16 = vqrshrn_n_s32(q13s32, 14);
+    d9s16 = vqrshrn_n_s32(q6s32, 14);
+    q3s16 = vcombine_s16(d6s16, d7s16);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+
+    // stage 7
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q15s16);
+    q13s16 = vaddq_s16(q1s16, q14s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q14s16 = vsubq_s16(q1s16, q14s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q10s16, q5s16);
+    q13s16 = vaddq_s16(q11s16, q4s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q4s16 = vsubq_s16(q11s16, q4s16);
+    q5s16 = vsubq_s16(q10s16, q5s16);
+
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q3s16);
+    q13s16 = vaddq_s16(q1s16, q2s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q2s16 = vsubq_s16(q1s16, q2s16);
+    q3s16 = vsubq_s16(q0s16, q3s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    q12s16 = vaddq_s16(q10s16, q9s16);
+    q13s16 = vaddq_s16(q11s16, q8s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q8s16 = vsubq_s16(q11s16, q8s16);
+    q9s16 = vsubq_s16(q10s16, q9s16);
+
+    d4u64  = vreinterpret_u64_s16(vget_low_s16(q2s16));
+    d5u64  = vreinterpret_u64_s16(vget_high_s16(q2s16));
+    d6u64  = vreinterpret_u64_s16(vget_low_s16(q3s16));
+    d7u64  = vreinterpret_u64_s16(vget_high_s16(q3s16));
+    d8u64  = vreinterpret_u64_s16(vget_low_s16(q4s16));
+    d9u64  = vreinterpret_u64_s16(vget_high_s16(q4s16));
+    d10u64 = vreinterpret_u64_s16(vget_low_s16(q5s16));
+    d11u64 = vreinterpret_u64_s16(vget_high_s16(q5s16));
+    d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+    d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+    d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+    d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+    d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+    d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+    d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+    d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+    vst1_u64((uint64_t *)out, d16u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d17u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d18u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d19u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d4u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d5u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d6u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d7u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d8u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d9u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d10u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d11u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d28u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d29u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d30u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d31u64);
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon_asm.asm
new file mode 100644
index 000000000..a13c0d04b
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_add_neon_asm.asm
@@ -0,0 +1,1179 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_idct16x16_256_add_neon_pass1|
+    EXPORT  |vp9_idct16x16_256_add_neon_pass2|
+    EXPORT  |vp9_idct16x16_10_add_neon_pass1|
+    EXPORT  |vp9_idct16x16_10_add_neon_pass2|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void |vp9_idct16x16_256_add_neon_pass1|(int16_t *input,
+;                                          int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vp9_idct16x16_256_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0xc00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r12, #0x3e00
+    add             r12, #0xc5
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r12                   ; duplicate cospi_4_64
+
+    ; preloading to avoid stall
+    ; generate cospi_12_64 = 13623
+    mov             r3, #0x3500
+    add             r3, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r12, #0x2300
+    add             r12, #0x8e
+
+    ; step2[4] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; step2[4] * cospi_4_64
+    vmull.s16       q5, d18, d1
+    vmull.s16       q6, d19, d1
+
+    ; temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64
+    vmlal.s16       q5, d30, d0
+    vmlal.s16       q6, d31, d0
+
+    vdup.16         d2, r3                    ; duplicate cospi_12_64
+    vdup.16         d3, r12                   ; duplicate cospi_20_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q5, #14              ; >> 14
+    vqrshrn.s32     d15, q6, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; step2[5] * cospi_12_64
+    vmull.s16       q2, d26, d2
+    vmull.s16       q3, d27, d2
+
+    ; step2[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q15, d27, d3
+
+    ; temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q2, d22, d3
+    vmlsl.s16       q3, d23, d3
+
+    ; temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q15, d23, d2
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q2, #14              ; >> 14
+    vqrshrn.s32     d11, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q15, #14             ; >> 14
+
+    ; stage 4
+    vdup.16         d30, r3                   ; cospi_16_64
+
+    ; step1[0] * cospi_16_64
+    vmull.s16       q2, d16, d30
+    vmull.s16       q11, d17, d30
+
+    ; step1[1] * cospi_16_64
+    vmull.s16       q0, d24, d30
+    vmull.s16       q1, d25, d30
+
+    ; generate cospi_8_64 = 15137
+    mov             r3, #0x3b00
+    add             r3, #0x21
+
+    vdup.16         d30, r12                  ; duplicate cospi_24_64
+    vdup.16         d31, r3                   ; duplicate cospi_8_64
+
+    ; temp1 = (step1[0] + step1[1]) * cospi_16_64
+    vadd.s32        q3, q2, q0
+    vadd.s32        q12, q11, q1
+
+    ; temp2 = (step1[0] - step1[1]) * cospi_16_64
+    vsub.s32        q13, q2, q0
+    vsub.s32        q1, q11, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d16, q3, #14              ; >> 14
+    vqrshrn.s32     d17, q12, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d18, q13, #14             ; >> 14
+    vqrshrn.s32     d19, q1, #14              ; >> 14
+
+    ; step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+    ; step1[2] * cospi_8_64
+    vmull.s16       q0, d20, d31
+    vmull.s16       q1, d21, d31
+
+    ; step1[2] * cospi_24_64
+    vmull.s16       q12, d20, d30
+    vmull.s16       q13, d21, d30
+
+    ; temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q0, d28, d30
+    vmlal.s16       q1, d29, d30
+
+    ; temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q12, d28, d31
+    vmlsl.s16       q13, d29, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d22, q0, #14              ; >> 14
+    vqrshrn.s32     d23, q1, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d20, q12, #14             ; >> 14
+    vqrshrn.s32     d21, q13, #14             ; >> 14
+
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5];
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5];
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7];
+    vadd.s16        q15, q6, q7               ; step2[7] = step1[6] + step1[7];
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; stage 5
+    vadd.s16        q0, q8, q11               ; step1[0] = step2[0] + step2[3];
+    vadd.s16        q1, q9, q10               ; step1[1] = step2[1] + step2[2];
+    vsub.s16        q2, q9, q10               ; step1[2] = step2[1] - step2[2];
+    vsub.s16        q3, q8, q11               ; step1[3] = step2[0] - step2[3];
+
+    vdup.16         d16, r3;                  ; duplicate cospi_16_64
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q11, d26, d16
+    vmull.s16       q12, d27, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q6, q9, q11
+    vsub.s32        q13, q10, q12
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+    vqrshrn.s32     d11, q13, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q8, q0, q15                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; step2[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; step2[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q15              ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d16}, [r1], r2
+    vst1.64         {d17}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |vp9_idct16x16_256_add_neon_pass1|
+
+;void vp9_idct16x16_256_add_neon_pass2(int16_t *src,
+;                                        int16_t *output,
+;                                        int16_t *pass1Output,
+;                                        int16_t skip_adding,
+;                                        uint8_t *dest,
+;                                        int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vp9_idct16x16_256_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate  cospi_30_64 = 1606
+    mov             r3, #0x0600
+    add             r3, #0x46
+
+    ; generate cospi_2_64  = 16305
+    mov             r12, #0x3f00
+    add             r12, #0xb1
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d12, r3                   ; duplicate cospi_30_64
+    vdup.16         d13, r12                  ; duplicate cospi_2_64
+
+    ; preloading to avoid stall
+    ; generate cospi_14_64 = 12665
+    mov             r3, #0x3100
+    add             r3, #0x79
+
+    ; generate cospi_18_64 = 10394
+    mov             r12, #0x2800
+    add             r12, #0x9a
+
+    ; step1[8] * cospi_30_64
+    vmull.s16       q2, d16, d12
+    vmull.s16       q3, d17, d12
+
+    ; step1[8] * cospi_2_64
+    vmull.s16       q1, d16, d13
+    vmull.s16       q4, d17, d13
+
+    ; temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64
+    vmlsl.s16       q2, d30, d13
+    vmlsl.s16       q3, d31, d13
+
+    ; temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64
+    vmlal.s16       q1, d30, d12
+    vmlal.s16       q4, d31, d12
+
+    vdup.16         d30, r3                   ; duplicate cospi_14_64
+    vdup.16         d31, r12                  ; duplicate cospi_18_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d0, q2, #14               ; >> 14
+    vqrshrn.s32     d1, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q1, #14              ; >> 14
+    vqrshrn.s32     d15, q4, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_22_64 = 7723
+    mov             r3, #0x1e00
+    add             r3, #0x2b
+
+    ; generate cospi_10_64 = 14449
+    mov             r12, #0x3800
+    add             r12, #0x71
+
+    ; step1[9] * cospi_14_64
+    vmull.s16       q2, d24, d30
+    vmull.s16       q3, d25, d30
+
+    ; step1[9] * cospi_18_64
+    vmull.s16       q4, d24, d31
+    vmull.s16       q5, d25, d31
+
+    ; temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64
+    vmlsl.s16       q2, d22, d31
+    vmlsl.s16       q3, d23, d31
+
+    ; temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64
+    vmlal.s16       q4, d22, d30
+    vmlal.s16       q5, d23, d30
+
+    vdup.16         d30, r3                   ; duplicate cospi_22_64
+    vdup.16         d31, r12                  ; duplicate cospi_10_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q2, #14               ; >> 14
+    vqrshrn.s32     d3, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q4, #14              ; >> 14
+    vqrshrn.s32     d13, q5, #14              ; >> 14
+
+    ; step1[10] * cospi_22_64
+    vmull.s16       q11, d20, d30
+    vmull.s16       q12, d21, d30
+
+    ; step1[10] * cospi_10_64
+    vmull.s16       q4, d20, d31
+    vmull.s16       q5, d21, d31
+
+    ; temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64
+    vmlsl.s16       q11, d26, d31
+    vmlsl.s16       q12, d27, d31
+
+    ; temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64
+    vmlal.s16       q4, d26, d30
+    vmlal.s16       q5, d27, d30
+
+    ; preloading to avoid stall
+    ; generate cospi_6_64 = 15679
+    mov             r3, #0x3d00
+    add             r3, #0x3f
+
+    ; generate cospi_26_64 = 4756
+    mov             r12, #0x1200
+    add             r12, #0x94
+
+    vdup.16         d30, r3                   ; duplicate cospi_6_64
+    vdup.16         d31, r12                  ; duplicate cospi_26_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d11, q5, #14              ; >> 14
+    vqrshrn.s32     d10, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_6_64
+    vmull.s16       q10, d28, d30
+    vmull.s16       q11, d29, d30
+
+    ; step1[11] * cospi_26_64
+    vmull.s16       q12, d28, d31
+    vmull.s16       q13, d29, d31
+
+    ; temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64
+    vmlsl.s16       q10, d18, d31
+    vmlsl.s16       q11, d19, d31
+
+    ; temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64
+    vmlal.s16       q12, d18, d30
+    vmlal.s16       q13, d19, d30
+
+    vsub.s16        q9, q0, q1                ; step1[9]=step2[8]-step2[9]
+    vadd.s16        q0, q0, q1                ; step1[8]=step2[8]+step2[9]
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q11, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q12, #14              ; >> 14
+    vqrshrn.s32     d9, q13, #14              ; >> 14
+
+    ; stage 3
+    vsub.s16        q10, q3, q2               ; step1[10]=-step2[10]+step2[11]
+    vadd.s16        q11, q2, q3               ; step1[11]=step2[10]+step2[11]
+    vadd.s16        q12, q4, q5               ; step1[12]=step2[12]+step2[13]
+    vsub.s16        q13, q4, q5               ; step1[13]=step2[12]-step2[13]
+    vsub.s16        q14, q7, q6               ; step1[14]=-step2[14]+tep2[15]
+    vadd.s16        q7, q6, q7                ; step1[15]=step2[14]+step2[15]
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+
+    ; -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d18, d31
+    vmull.s16       q3, d19, d31
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q4, d28, d31
+    vmull.s16       q5, d29, d31
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d28, d30
+    vmlal.s16       q3, d29, d30
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q4, d18, d30
+    vmlsl.s16       q5, d19, d30
+
+    rsb             r12, #0
+    vdup.16         d30, r12                  ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q4, #14               ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    vmov.s16        q3, q11
+    vmov.s16        q4, q12
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q11, d26, d30
+    vmull.s16       q12, d27, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d20, d30
+    vmull.s16       q9, d21, d30
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlsl.s16       q11, d20, d31
+    vmlsl.s16       q12, d21, d31
+
+    ; temp1 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d26, d31
+    vmlal.s16       q9, d27, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q10, q3, q0
+    vadd.s32        q4, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q10, #14             ; >> 14
+    vqrshrn.s32     d11, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+    cmp              r3, #0                   ; check if need adding dest data
+    beq              skip_adding_dest
+
+    ldr              r7, [sp, #28]            ; dest used to save element 0-7
+    mov              r9, r7                   ; save dest pointer for later use
+    ldr              r8, [sp, #32]            ; load dest_stride
+
+    ; stage 7
+    ; load the data in pass1
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vrshr.s16       q8, q8, #6                ; ROUND_POWER_OF_TWO
+    vaddw.u8        q8, q8, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q8                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q9, q9, #6
+    vaddw.u8        q9, q9, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q9                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q2, q2, #6
+    vaddw.u8        q2, q2, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q2                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q3, q3, #6
+    vaddw.u8        q3, q3, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q3                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q4, q4, #6
+    vaddw.u8        q4, q4, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q4                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q5, q5, #6
+    vaddw.u8        q5, q5, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q5                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q14, q14, #6
+    vaddw.u8        q14, q14, d12             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q14                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q15, q15, #6
+    vaddw.u8        q15, q15, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q15                  ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    b               end_idct16x16_pass2
+
+skip_adding_dest
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |vp9_idct16x16_256_add_neon_pass2|
+
+;void |vp9_idct16x16_10_add_neon_pass1|(int16_t *input,
+;                                             int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vp9_idct16x16_10_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64*2 = 6392
+    mov             r3, #0x1800
+    add             r3, #0xf8
+
+    ; generate cospi_4_64*2  = 32138
+    mov             r12, #0x7d00
+    add             r12, #0x8a
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q0, r3                    ; duplicate cospi_28_64*2
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(step2[4] * cospi_28_64). vvqrdmulh will multiply,
+    ; double, and return the high 16 bits, effectively giving >> 15. Doubling
+    ; the constant will change this to >> 14.
+    ; dct_const_round_shift(step2[4] * cospi_28_64);
+    vqrdmulh.s16    q4, q9, q0
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64*2 = 23170
+    mov             r3, #0x5a00
+    add             r3, #0x82
+
+    ; dct_const_round_shift(step2[4] * cospi_4_64);
+    vqrdmulh.s16    q7, q9, q1
+
+    ; stage 4
+    vdup.16         q1, r3                    ; cospi_16_64*2
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    vdup.16         d4, r3;                   ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(step1[0] * cospi_16_64)
+    vqrdmulh.s16    q8, q8, q1
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d14, d4
+    vmull.s16       q10, d15, d4
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q12, d9, d4
+    vmull.s16       q11, d8, d4
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q15, q10, q12
+    vsub.s32        q6, q9, q11
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d11, q15, #14             ; >> 14
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q2, q8, q7                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q10, q8, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q8, q4               ; step2[3] = step1[3] + step1[4];
+    vadd.s16        q9, q8, q6                ; step2[1] = step1[1] + step1[6];
+    vsub.s16        q12, q8, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q8, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q8, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q8, q7               ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d4}, [r1], r2
+    vst1.64         {d5}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |vp9_idct16x16_10_add_neon_pass1|
+
+;void vp9_idct16x16_10_add_neon_pass2(int16_t *src,
+;                                           int16_t *output,
+;                                           int16_t *pass1Output,
+;                                           int16_t skip_adding,
+;                                           uint8_t *dest,
+;                                           int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|vp9_idct16x16_10_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate 2*cospi_30_64 = 3212
+    mov             r3, #0xc00
+    add             r3, #0x8c
+
+    ; generate 2*cospi_2_64  = 32610
+    mov             r12, #0x7f00
+    add             r12, #0x62
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q6, r3                    ; duplicate 2*cospi_30_64
+
+    ; dct_const_round_shift(step1[8] * cospi_30_64)
+    vqrdmulh.s16    q0, q8, q6
+
+    vdup.16         q6, r12                   ; duplicate 2*cospi_2_64
+
+    ; dct_const_round_shift(step1[8] * cospi_2_64)
+    vqrdmulh.s16    q7, q8, q6
+
+    ; preloading to avoid stall
+    ; generate 2*cospi_26_64 = 9512
+    mov             r12, #0x2500
+    add             r12, #0x28
+    rsb             r12, #0
+    vdup.16         q15, r12                  ; duplicate -2*cospi_26_64
+
+    ; generate 2*cospi_6_64 = 31358
+    mov             r3, #0x7a00
+    add             r3, #0x7e
+    vdup.16         q14, r3                   ; duplicate 2*cospi_6_64
+
+    ; dct_const_round_shift(- step1[12] * cospi_26_64)
+    vqrdmulh.s16    q3, q9, q15
+
+    ; dct_const_round_shift(step1[12] * cospi_6_64)
+    vqrdmulh.s16    q4, q9, q14
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q12, d14, d31
+    vmull.s16       q5, d15, d31
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d0, d31
+    vmull.s16       q11, d1, d31
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q12, d0, d30
+    vmlsl.s16       q5, d1, d30
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d14, d30
+    vmlal.s16       q11, d15, d30
+
+    rsb              r12, #0
+    vdup.16          d30, r12                 ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q12, #14              ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q11, #14             ; >> 14
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q10, d8, d30
+    vmull.s16       q13, d9, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d6, d30
+    vmull.s16       q9, d7, d30
+
+    ; temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64
+    vmlsl.s16       q10, d6, d31
+    vmlsl.s16       q13, d7, d31
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d8, d31
+    vmlal.s16       q9, d9, d31
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q10, #14              ; >> 14
+    vqrshrn.s32     d5, q13, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q0, q3, q0
+    vadd.s32        q1, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q0, #14              ; >> 14
+    vqrshrn.s32     d11, q1, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift((step1[11] + step1[12]) * cospi_16_64);
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct10_16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |vp9_idct16x16_10_add_neon_pass2|
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_neon.c
new file mode 100644
index 000000000..f2c4ec451
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct16x16_neon.c
@@ -0,0 +1,186 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+
+void vp9_idct16x16_256_add_neon_pass1(const int16_t *input,
+                                      int16_t *output,
+                                      int output_stride);
+void vp9_idct16x16_256_add_neon_pass2(const int16_t *src,
+                                      int16_t *output,
+                                      int16_t *pass1Output,
+                                      int16_t skip_adding,
+                                      uint8_t *dest,
+                                      int dest_stride);
+void vp9_idct16x16_10_add_neon_pass1(const int16_t *input,
+                                     int16_t *output,
+                                     int output_stride);
+void vp9_idct16x16_10_add_neon_pass2(const int16_t *src,
+                                     int16_t *output,
+                                     int16_t *pass1Output,
+                                     int16_t skip_adding,
+                                     uint8_t *dest,
+                                     int dest_stride);
+
+#if HAVE_NEON_ASM
+/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */
+extern void vp9_push_neon(int64_t *store);
+extern void vp9_pop_neon(int64_t *store);
+#endif  // HAVE_NEON_ASM
+
+void vp9_idct16x16_256_add_neon(const int16_t *input,
+                                uint8_t *dest, int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16*16] = {0};
+  int16_t row_idct_output[16*16] = {0};
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  vp9_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vp9_idct16x16_256_add_neon_pass2(input+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     0,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the lower 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(input+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vp9_idct16x16_256_add_neon_pass2(input+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     0,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vp9_idct16x16_256_add_neon_pass2(row_idct_output+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     1,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(row_idct_output+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vp9_idct16x16_256_add_neon_pass2(row_idct_output+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     1,
+                                     dest+8,
+                                     dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  vp9_pop_neon(store_reg);
+#endif
+
+  return;
+}
+
+void vp9_idct16x16_10_add_neon(const int16_t *input,
+                               uint8_t *dest, int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16*16] = {0};
+  int16_t row_idct_output[16*16] = {0};
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  vp9_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_10_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  vp9_idct16x16_10_add_neon_pass2(input+1,
+                                        row_idct_output,
+                                        pass1_output,
+                                        0,
+                                        dest,
+                                        dest_stride);
+
+  /* Skip Parallel idct on the lower 8 rows as they are all 0s */
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vp9_idct16x16_256_add_neon_pass2(row_idct_output+1,
+                                     row_idct_output,
+                                     pass1_output,
+                                     1,
+                                     dest,
+                                     dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  vp9_idct16x16_256_add_neon_pass1(row_idct_output+8*16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  vp9_idct16x16_256_add_neon_pass2(row_idct_output+8*16+1,
+                                     row_idct_output+8,
+                                     pass1_output,
+                                     1,
+                                     dest+8,
+                                     dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  vp9_pop_neon(store_reg);
+#endif
+
+  return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.c
new file mode 100644
index 000000000..0ce45f2bf
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon.c
@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_idct.h"
+
+static INLINE void LD_16x8(
+        uint8_t *d,
+        int d_stride,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vld1q_u8(d);
+    d += d_stride;
+    *q9u8 = vld1q_u8(d);
+    d += d_stride;
+    *q10u8 = vld1q_u8(d);
+    d += d_stride;
+    *q11u8 = vld1q_u8(d);
+    d += d_stride;
+    *q12u8 = vld1q_u8(d);
+    d += d_stride;
+    *q13u8 = vld1q_u8(d);
+    d += d_stride;
+    *q14u8 = vld1q_u8(d);
+    d += d_stride;
+    *q15u8 = vld1q_u8(d);
+    return;
+}
+
+static INLINE void ADD_DIFF_16x8(
+        uint8x16_t qdiffu8,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vqaddq_u8(*q8u8, qdiffu8);
+    *q9u8 = vqaddq_u8(*q9u8, qdiffu8);
+    *q10u8 = vqaddq_u8(*q10u8, qdiffu8);
+    *q11u8 = vqaddq_u8(*q11u8, qdiffu8);
+    *q12u8 = vqaddq_u8(*q12u8, qdiffu8);
+    *q13u8 = vqaddq_u8(*q13u8, qdiffu8);
+    *q14u8 = vqaddq_u8(*q14u8, qdiffu8);
+    *q15u8 = vqaddq_u8(*q15u8, qdiffu8);
+    return;
+}
+
+static INLINE void SUB_DIFF_16x8(
+        uint8x16_t qdiffu8,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    *q8u8 = vqsubq_u8(*q8u8, qdiffu8);
+    *q9u8 = vqsubq_u8(*q9u8, qdiffu8);
+    *q10u8 = vqsubq_u8(*q10u8, qdiffu8);
+    *q11u8 = vqsubq_u8(*q11u8, qdiffu8);
+    *q12u8 = vqsubq_u8(*q12u8, qdiffu8);
+    *q13u8 = vqsubq_u8(*q13u8, qdiffu8);
+    *q14u8 = vqsubq_u8(*q14u8, qdiffu8);
+    *q15u8 = vqsubq_u8(*q15u8, qdiffu8);
+    return;
+}
+
+static INLINE void ST_16x8(
+        uint8_t *d,
+        int d_stride,
+        uint8x16_t *q8u8,
+        uint8x16_t *q9u8,
+        uint8x16_t *q10u8,
+        uint8x16_t *q11u8,
+        uint8x16_t *q12u8,
+        uint8x16_t *q13u8,
+        uint8x16_t *q14u8,
+        uint8x16_t *q15u8) {
+    vst1q_u8(d, *q8u8);
+    d += d_stride;
+    vst1q_u8(d, *q9u8);
+    d += d_stride;
+    vst1q_u8(d, *q10u8);
+    d += d_stride;
+    vst1q_u8(d, *q11u8);
+    d += d_stride;
+    vst1q_u8(d, *q12u8);
+    d += d_stride;
+    vst1q_u8(d, *q13u8);
+    d += d_stride;
+    vst1q_u8(d, *q14u8);
+    d += d_stride;
+    vst1q_u8(d, *q15u8);
+    return;
+}
+
+void vp9_idct32x32_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x16_t q0u8, q8u8, q9u8, q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int i, j, dest_stride8;
+    uint8_t *d;
+    int16_t a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 6);
+
+    dest_stride8 = dest_stride * 8;
+    if (a1 >= 0) {  // diff_positive_32_32
+        a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+        q0u8 = vdupq_n_u8(a1);
+        for (i = 0; i < 2; i++, dest += 16) {  // diff_positive_32_32_loop
+            d = dest;
+            for (j = 0; j < 4; j++) {
+                LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                ADD_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8,
+                                    &q12u8, &q13u8, &q14u8, &q15u8);
+                ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                d += dest_stride8;
+            }
+        }
+    } else {  // diff_negative_32_32
+        a1 = -a1;
+        a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+        q0u8 = vdupq_n_u8(a1);
+        for (i = 0; i < 2; i++, dest += 16) {  // diff_negative_32_32_loop
+            d = dest;
+            for (j = 0; j < 4; j++) {
+                LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                SUB_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8,
+                                    &q12u8, &q13u8, &q14u8, &q15u8);
+                ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8,
+                                        &q12u8, &q13u8, &q14u8, &q15u8);
+                d += dest_stride8;
+            }
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon_asm.asm
new file mode 100644
index 000000000..d290d0753
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_1_add_neon_asm.asm
@@ -0,0 +1,144 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+    EXPORT  |vp9_idct32x32_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ;TODO(hkuang): put the following macros in a seperate
+    ;file so other idct function could also use them.
+    MACRO
+    LD_16x8          $src, $stride
+    vld1.8           {q8}, [$src], $stride
+    vld1.8           {q9}, [$src], $stride
+    vld1.8           {q10}, [$src], $stride
+    vld1.8           {q11}, [$src], $stride
+    vld1.8           {q12}, [$src], $stride
+    vld1.8           {q13}, [$src], $stride
+    vld1.8           {q14}, [$src], $stride
+    vld1.8           {q15}, [$src], $stride
+    MEND
+
+    MACRO
+    ADD_DIFF_16x8    $diff
+    vqadd.u8         q8, q8, $diff
+    vqadd.u8         q9, q9, $diff
+    vqadd.u8         q10, q10, $diff
+    vqadd.u8         q11, q11, $diff
+    vqadd.u8         q12, q12, $diff
+    vqadd.u8         q13, q13, $diff
+    vqadd.u8         q14, q14, $diff
+    vqadd.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    SUB_DIFF_16x8    $diff
+    vqsub.u8         q8, q8, $diff
+    vqsub.u8         q9, q9, $diff
+    vqsub.u8         q10, q10, $diff
+    vqsub.u8         q11, q11, $diff
+    vqsub.u8         q12, q12, $diff
+    vqsub.u8         q13, q13, $diff
+    vqsub.u8         q14, q14, $diff
+    vqsub.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    ST_16x8          $dst, $stride
+    vst1.8           {q8}, [$dst], $stride
+    vst1.8           {q9}, [$dst], $stride
+    vst1.8           {q10},[$dst], $stride
+    vst1.8           {q11},[$dst], $stride
+    vst1.8           {q12},[$dst], $stride
+    vst1.8           {q13},[$dst], $stride
+    vst1.8           {q14},[$dst], $stride
+    vst1.8           {q15},[$dst], $stride
+    MEND
+
+;void vp9_idct32x32_1_add_neon(int16_t *input, uint8_t *dest,
+;                              int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride
+
+|vp9_idct32x32_1_add_neon| PROC
+    push             {lr}
+    pld              [r1]
+    add              r3, r1, #16               ; r3 dest + 16 for second loop
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asrs             r0, r0, #6                ; >> 6
+    bge              diff_positive_32_32
+
+diff_negative_32_32
+    neg              r0, r0
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_negative_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_negative_32_32_loop
+    pop              {pc}
+
+diff_positive_32_32
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_positive_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_positive_32_32_loop
+    pop              {pc}
+
+    ENDP             ; |vp9_idct32x32_1_add_neon|
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.c
new file mode 100644
index 000000000..309bdf8d7
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon.c
@@ -0,0 +1,750 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+static int16_t cospi_1_64 = 16364;
+static int16_t cospi_2_64 = 16305;
+static int16_t cospi_3_64 = 16207;
+static int16_t cospi_4_64 = 16069;
+static int16_t cospi_5_64 = 15893;
+static int16_t cospi_6_64 = 15679;
+static int16_t cospi_7_64 = 15426;
+static int16_t cospi_8_64 = 15137;
+static int16_t cospi_9_64 = 14811;
+static int16_t cospi_10_64 = 14449;
+static int16_t cospi_11_64 = 14053;
+static int16_t cospi_12_64 = 13623;
+static int16_t cospi_13_64 = 13160;
+static int16_t cospi_14_64 = 12665;
+static int16_t cospi_15_64 = 12140;
+static int16_t cospi_16_64 = 11585;
+static int16_t cospi_17_64 = 11003;
+static int16_t cospi_18_64 = 10394;
+static int16_t cospi_19_64 = 9760;
+static int16_t cospi_20_64 = 9102;
+static int16_t cospi_21_64 = 8423;
+static int16_t cospi_22_64 = 7723;
+static int16_t cospi_23_64 = 7005;
+static int16_t cospi_24_64 = 6270;
+static int16_t cospi_25_64 = 5520;
+static int16_t cospi_26_64 = 4756;
+static int16_t cospi_27_64 = 3981;
+static int16_t cospi_28_64 = 3196;
+static int16_t cospi_29_64 = 2404;
+static int16_t cospi_30_64 = 1606;
+static int16_t cospi_31_64 = 804;
+
+#define LOAD_FROM_TRANSPOSED(prev, first, second) \
+    q14s16 = vld1q_s16(trans_buf + first * 8); \
+    q13s16 = vld1q_s16(trans_buf + second * 8);
+
+#define LOAD_FROM_OUTPUT(prev, first, second, qA, qB) \
+    qA = vld1q_s16(out + first * 32); \
+    qB = vld1q_s16(out + second * 32);
+
+#define STORE_IN_OUTPUT(prev, first, second, qA, qB) \
+    vst1q_s16(out + first * 32, qA); \
+    vst1q_s16(out + second * 32, qB);
+
+#define  STORE_COMBINE_CENTER_RESULTS(r10, r9) \
+       __STORE_COMBINE_CENTER_RESULTS(r10, r9, stride, \
+                                      q6s16, q7s16, q8s16, q9s16);
+static INLINE void __STORE_COMBINE_CENTER_RESULTS(
+        uint8_t *p1,
+        uint8_t *p2,
+        int stride,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16) {
+    int16x4_t d8s16, d9s16, d10s16, d11s16;
+
+    d8s16 = vld1_s16((int16_t *)p1);
+    p1 += stride;
+    d11s16 = vld1_s16((int16_t *)p2);
+    p2 -= stride;
+    d9s16 = vld1_s16((int16_t *)p1);
+    d10s16 = vld1_s16((int16_t *)p2);
+
+    q7s16 = vrshrq_n_s16(q7s16, 6);
+    q8s16 = vrshrq_n_s16(q8s16, 6);
+    q9s16 = vrshrq_n_s16(q9s16, 6);
+    q6s16 = vrshrq_n_s16(q6s16, 6);
+
+    q7s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q7s16),
+                                           vreinterpret_u8_s16(d9s16)));
+    q8s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                                           vreinterpret_u8_s16(d10s16)));
+    q9s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                                           vreinterpret_u8_s16(d11s16)));
+    q6s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q6s16),
+                                           vreinterpret_u8_s16(d8s16)));
+
+    d9s16  = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+    d10s16 = vreinterpret_s16_u8(vqmovun_s16(q8s16));
+    d11s16 = vreinterpret_s16_u8(vqmovun_s16(q9s16));
+    d8s16  = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+
+    vst1_s16((int16_t *)p1, d9s16);
+    p1 -= stride;
+    vst1_s16((int16_t *)p2, d10s16);
+    p2 += stride;
+    vst1_s16((int16_t *)p1, d8s16);
+    vst1_s16((int16_t *)p2, d11s16);
+    return;
+}
+
+#define  STORE_COMBINE_EXTREME_RESULTS(r7, r6); \
+       __STORE_COMBINE_EXTREME_RESULTS(r7, r6, stride, \
+                                      q4s16, q5s16, q6s16, q7s16);
+static INLINE void __STORE_COMBINE_EXTREME_RESULTS(
+        uint8_t *p1,
+        uint8_t *p2,
+        int stride,
+        int16x8_t q4s16,
+        int16x8_t q5s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16) {
+    int16x4_t d4s16, d5s16, d6s16, d7s16;
+
+    d4s16 = vld1_s16((int16_t *)p1);
+    p1 += stride;
+    d7s16 = vld1_s16((int16_t *)p2);
+    p2 -= stride;
+    d5s16 = vld1_s16((int16_t *)p1);
+    d6s16 = vld1_s16((int16_t *)p2);
+
+    q5s16 = vrshrq_n_s16(q5s16, 6);
+    q6s16 = vrshrq_n_s16(q6s16, 6);
+    q7s16 = vrshrq_n_s16(q7s16, 6);
+    q4s16 = vrshrq_n_s16(q4s16, 6);
+
+    q5s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q5s16),
+                                           vreinterpret_u8_s16(d5s16)));
+    q6s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q6s16),
+                                           vreinterpret_u8_s16(d6s16)));
+    q7s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q7s16),
+                                           vreinterpret_u8_s16(d7s16)));
+    q4s16 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q4s16),
+                                           vreinterpret_u8_s16(d4s16)));
+
+    d5s16 = vreinterpret_s16_u8(vqmovun_s16(q5s16));
+    d6s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+    d7s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+    d4s16 = vreinterpret_s16_u8(vqmovun_s16(q4s16));
+
+    vst1_s16((int16_t *)p1, d5s16);
+    p1 -= stride;
+    vst1_s16((int16_t *)p2, d6s16);
+    p2 += stride;
+    vst1_s16((int16_t *)p2, d7s16);
+    vst1_s16((int16_t *)p1, d4s16);
+    return;
+}
+
+#define DO_BUTTERFLY_STD(const_1, const_2, qA, qB) \
+        DO_BUTTERFLY(q14s16, q13s16, const_1, const_2, qA, qB);
+static INLINE void DO_BUTTERFLY(
+        int16x8_t q14s16,
+        int16x8_t q13s16,
+        int16_t first_const,
+        int16_t second_const,
+        int16x8_t *qAs16,
+        int16x8_t *qBs16) {
+    int16x4_t d30s16, d31s16;
+    int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q15s32;
+    int16x4_t dCs16, dDs16, dAs16, dBs16;
+
+    dCs16 = vget_low_s16(q14s16);
+    dDs16 = vget_high_s16(q14s16);
+    dAs16 = vget_low_s16(q13s16);
+    dBs16 = vget_high_s16(q13s16);
+
+    d30s16 = vdup_n_s16(first_const);
+    d31s16 = vdup_n_s16(second_const);
+
+    q8s32 = vmull_s16(dCs16, d30s16);
+    q10s32 = vmull_s16(dAs16, d31s16);
+    q9s32 = vmull_s16(dDs16, d30s16);
+    q11s32 = vmull_s16(dBs16, d31s16);
+    q12s32 = vmull_s16(dCs16, d31s16);
+
+    q8s32 = vsubq_s32(q8s32, q10s32);
+    q9s32 = vsubq_s32(q9s32, q11s32);
+
+    q10s32 = vmull_s16(dDs16, d31s16);
+    q11s32 = vmull_s16(dAs16, d30s16);
+    q15s32 = vmull_s16(dBs16, d30s16);
+
+    q11s32 = vaddq_s32(q12s32, q11s32);
+    q10s32 = vaddq_s32(q10s32, q15s32);
+
+    *qAs16 = vcombine_s16(vqrshrn_n_s32(q8s32, 14),
+                          vqrshrn_n_s32(q9s32, 14));
+    *qBs16 = vcombine_s16(vqrshrn_n_s32(q11s32, 14),
+                          vqrshrn_n_s32(q10s32, 14));
+    return;
+}
+
+static INLINE void idct32_transpose_pair(
+        int16_t *input,
+        int16_t *t_buf) {
+    int16_t *in;
+    int i;
+    const int stride = 32;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    for (i = 0; i < 4; i++, input += 8) {
+        in = input;
+        q8s16 = vld1q_s16(in);
+        in += stride;
+        q9s16 = vld1q_s16(in);
+        in += stride;
+        q10s16 = vld1q_s16(in);
+        in += stride;
+        q11s16 = vld1q_s16(in);
+        in += stride;
+        q12s16 = vld1q_s16(in);
+        in += stride;
+        q13s16 = vld1q_s16(in);
+        in += stride;
+        q14s16 = vld1q_s16(in);
+        in += stride;
+        q15s16 = vld1q_s16(in);
+
+        d16s16 = vget_low_s16(q8s16);
+        d17s16 = vget_high_s16(q8s16);
+        d18s16 = vget_low_s16(q9s16);
+        d19s16 = vget_high_s16(q9s16);
+        d20s16 = vget_low_s16(q10s16);
+        d21s16 = vget_high_s16(q10s16);
+        d22s16 = vget_low_s16(q11s16);
+        d23s16 = vget_high_s16(q11s16);
+        d24s16 = vget_low_s16(q12s16);
+        d25s16 = vget_high_s16(q12s16);
+        d26s16 = vget_low_s16(q13s16);
+        d27s16 = vget_high_s16(q13s16);
+        d28s16 = vget_low_s16(q14s16);
+        d29s16 = vget_high_s16(q14s16);
+        d30s16 = vget_low_s16(q15s16);
+        d31s16 = vget_high_s16(q15s16);
+
+        q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+        q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+        q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+        q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+        q12s16 = vcombine_s16(d17s16, d25s16);
+        q13s16 = vcombine_s16(d19s16, d27s16);
+        q14s16 = vcombine_s16(d21s16, d29s16);
+        q15s16 = vcombine_s16(d23s16, d31s16);
+
+        q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                            vreinterpretq_s32_s16(q10s16));
+        q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q9s16),
+                            vreinterpretq_s32_s16(q11s16));
+        q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q12s16),
+                            vreinterpretq_s32_s16(q14s16));
+        q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q13s16),
+                            vreinterpretq_s32_s16(q15s16));
+
+        q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                            vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+        q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                            vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+        q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                            vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+        q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                            vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+        vst1q_s16(t_buf, q0x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q0x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q1x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q1x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q2x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q2x2s16.val[1]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q3x2s16.val[0]);
+        t_buf += 8;
+        vst1q_s16(t_buf, q3x2s16.val[1]);
+        t_buf += 8;
+    }
+    return;
+}
+
+static INLINE void idct32_bands_end_1st_pass(
+        int16_t *out,
+        int16x8_t q2s16,
+        int16x8_t q3s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16,
+        int16x8_t q10s16,
+        int16x8_t q11s16,
+        int16x8_t q12s16,
+        int16x8_t q13s16,
+        int16x8_t q14s16,
+        int16x8_t q15s16) {
+    int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+    STORE_IN_OUTPUT(17, 16, 17, q6s16, q7s16);
+    STORE_IN_OUTPUT(17, 14, 15, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(15, 30, 31, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(31, 30, 31, q6s16, q7s16);
+    STORE_IN_OUTPUT(31, 0, 1, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(1, 12, 13, q0s16, q1s16);
+    q2s16 = vaddq_s16(q10s16, q1s16);
+    q3s16 = vaddq_s16(q11s16, q0s16);
+    q4s16 = vsubq_s16(q11s16, q0s16);
+    q5s16 = vsubq_s16(q10s16, q1s16);
+
+    LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(19, 18, 19, q6s16, q7s16);
+    STORE_IN_OUTPUT(19, 12, 13, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(13, 28, 29, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(29, 28, 29, q6s16, q7s16);
+    STORE_IN_OUTPUT(29, 2, 3, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(3, 10, 11, q0s16, q1s16);
+    q2s16 = vaddq_s16(q12s16, q1s16);
+    q3s16 = vaddq_s16(q13s16, q0s16);
+    q4s16 = vsubq_s16(q13s16, q0s16);
+    q5s16 = vsubq_s16(q12s16, q1s16);
+
+    LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(21, 20, 21, q6s16, q7s16);
+    STORE_IN_OUTPUT(21, 10, 11, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(11, 26, 27, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(27, 26, 27, q6s16, q7s16);
+    STORE_IN_OUTPUT(27, 4, 5, q4s16, q5s16);
+
+    LOAD_FROM_OUTPUT(5, 8, 9, q0s16, q1s16);
+    q2s16 = vaddq_s16(q14s16, q1s16);
+    q3s16 = vaddq_s16(q15s16, q0s16);
+    q4s16 = vsubq_s16(q15s16, q0s16);
+    q5s16 = vsubq_s16(q14s16, q1s16);
+
+    LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16);
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_IN_OUTPUT(23, 22, 23, q6s16, q7s16);
+    STORE_IN_OUTPUT(23, 8, 9, q8s16, q9s16);
+
+    LOAD_FROM_OUTPUT(9, 24, 25, q0s16, q1s16);
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_IN_OUTPUT(25, 24, 25, q6s16, q7s16);
+    STORE_IN_OUTPUT(25, 6, 7, q4s16, q5s16);
+    return;
+}
+
+static INLINE void idct32_bands_end_2nd_pass(
+        int16_t *out,
+        uint8_t *dest,
+        int stride,
+        int16x8_t q2s16,
+        int16x8_t q3s16,
+        int16x8_t q6s16,
+        int16x8_t q7s16,
+        int16x8_t q8s16,
+        int16x8_t q9s16,
+        int16x8_t q10s16,
+        int16x8_t q11s16,
+        int16x8_t q12s16,
+        int16x8_t q13s16,
+        int16x8_t q14s16,
+        int16x8_t q15s16) {
+    uint8_t *r6  = dest + 31 * stride;
+    uint8_t *r7  = dest/* +  0 * stride*/;
+    uint8_t *r9  = dest + 15 * stride;
+    uint8_t *r10 = dest + 16 * stride;
+    int str2 = stride << 1;
+    int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(17, 30, 31, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(31, 12, 13, q0s16, q1s16)
+    q2s16 = vaddq_s16(q10s16, q1s16);
+    q3s16 = vaddq_s16(q11s16, q0s16);
+    q4s16 = vsubq_s16(q11s16, q0s16);
+    q5s16 = vsubq_s16(q10s16, q1s16);
+
+    LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(19, 28, 29, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(29, 10, 11, q0s16, q1s16)
+    q2s16 = vaddq_s16(q12s16, q1s16);
+    q3s16 = vaddq_s16(q13s16, q0s16);
+    q4s16 = vsubq_s16(q13s16, q0s16);
+    q5s16 = vsubq_s16(q12s16, q1s16);
+
+    LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+    r10 += str2; r9 -= str2;
+
+    LOAD_FROM_OUTPUT(21, 26, 27, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    r7 += str2; r6 -= str2;
+
+    LOAD_FROM_OUTPUT(27, 8, 9, q0s16, q1s16)
+    q2s16 = vaddq_s16(q14s16, q1s16);
+    q3s16 = vaddq_s16(q15s16, q0s16);
+    q4s16 = vsubq_s16(q15s16, q0s16);
+    q5s16 = vsubq_s16(q14s16, q1s16);
+
+    LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16)
+    q8s16 = vaddq_s16(q4s16, q1s16);
+    q9s16 = vaddq_s16(q5s16, q0s16);
+    q6s16 = vsubq_s16(q5s16, q0s16);
+    q7s16 = vsubq_s16(q4s16, q1s16);
+    STORE_COMBINE_CENTER_RESULTS(r10, r9);
+
+    LOAD_FROM_OUTPUT(23, 24, 25, q0s16, q1s16)
+    q4s16 = vaddq_s16(q2s16, q1s16);
+    q5s16 = vaddq_s16(q3s16, q0s16);
+    q6s16 = vsubq_s16(q3s16, q0s16);
+    q7s16 = vsubq_s16(q2s16, q1s16);
+    STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+    return;
+}
+
+void vp9_idct32x32_1024_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int stride) {
+    int i, idct32_pass_loop;
+    int16_t trans_buf[32 * 8];
+    int16_t pass1[32 * 32];
+    int16_t pass2[32 * 32];
+    int16_t *out;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+
+    for (idct32_pass_loop = 0, out = pass1;
+         idct32_pass_loop < 2;
+         idct32_pass_loop++,
+         input = pass1,  // the input of pass2 is the result of pass1
+         out = pass2) {
+        for (i = 0;
+             i < 4; i++,
+             input += 32 * 8, out += 8) {  // idct32_bands_loop
+            idct32_transpose_pair(input, trans_buf);
+
+            // -----------------------------------------
+            // BLOCK A: 16-19,28-31
+            // -----------------------------------------
+            // generate 16,17,30,31
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(0, 1, 31)
+            DO_BUTTERFLY_STD(cospi_31_64, cospi_1_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(31, 17, 15)
+            DO_BUTTERFLY_STD(cospi_15_64, cospi_17_64, &q1s16, &q3s16)
+            // part of stage 2
+            q4s16 = vaddq_s16(q0s16, q1s16);
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q6s16 = vaddq_s16(q2s16, q3s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q5s16, &q7s16)
+
+            // generate 18,19,28,29
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(15, 9, 23)
+            DO_BUTTERFLY_STD(cospi_23_64, cospi_9_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(23, 25, 7)
+            DO_BUTTERFLY_STD(cospi_7_64, cospi_25_64, &q1s16, &q3s16)
+            // part of stage 2
+            q13s16 = vsubq_s16(q3s16, q2s16);
+            q3s16 = vaddq_s16(q3s16, q2s16);
+            q14s16 = vsubq_s16(q1s16, q0s16);
+            q2s16 = vaddq_s16(q1s16, q0s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(-cospi_4_64, -cospi_28_64, &q1s16, &q0s16)
+            // part of stage 4
+            q8s16 = vaddq_s16(q4s16, q2s16);
+            q9s16 = vaddq_s16(q5s16, q0s16);
+            q10s16 = vaddq_s16(q7s16, q1s16);
+            q15s16 = vaddq_s16(q6s16, q3s16);
+            q13s16 = vsubq_s16(q5s16, q0s16);
+            q14s16 = vsubq_s16(q7s16, q1s16);
+            STORE_IN_OUTPUT(0, 16, 31, q8s16, q15s16)
+            STORE_IN_OUTPUT(31, 17, 30, q9s16, q10s16)
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q0s16, &q1s16)
+            STORE_IN_OUTPUT(30, 29, 18, q1s16, q0s16)
+            // part of stage 4
+            q13s16 = vsubq_s16(q4s16, q2s16);
+            q14s16 = vsubq_s16(q6s16, q3s16);
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q4s16, &q6s16)
+            STORE_IN_OUTPUT(18, 19, 28, q4s16, q6s16)
+
+            // -----------------------------------------
+            // BLOCK B: 20-23,24-27
+            // -----------------------------------------
+            // generate 20,21,26,27
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(7, 5, 27)
+            DO_BUTTERFLY_STD(cospi_27_64, cospi_5_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(27, 21, 11)
+            DO_BUTTERFLY_STD(cospi_11_64, cospi_21_64, &q1s16, &q3s16)
+            // part of stage 2
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+
+            // generate 22,23,24,25
+            // part of stage 1
+            LOAD_FROM_TRANSPOSED(11, 13, 19)
+            DO_BUTTERFLY_STD(cospi_19_64, cospi_13_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(19, 29, 3)
+            DO_BUTTERFLY_STD(cospi_3_64, cospi_29_64, &q4s16, &q6s16)
+            // part of stage 2
+            q14s16 = vsubq_s16(q4s16, q5s16);
+            q5s16  = vaddq_s16(q4s16, q5s16);
+            q13s16 = vsubq_s16(q6s16, q7s16);
+            q6s16  = vaddq_s16(q6s16, q7s16);
+            // part of stage 3
+            DO_BUTTERFLY_STD(-cospi_20_64, -cospi_12_64, &q4s16, &q7s16)
+            // part of stage 4
+            q10s16 = vaddq_s16(q7s16, q1s16);
+            q11s16 = vaddq_s16(q5s16, q0s16);
+            q12s16 = vaddq_s16(q6s16, q2s16);
+            q15s16 = vaddq_s16(q4s16, q3s16);
+            // part of stage 6
+            LOAD_FROM_OUTPUT(28, 16, 17, q14s16, q13s16)
+            q8s16 = vaddq_s16(q14s16, q11s16);
+            q9s16 = vaddq_s16(q13s16, q10s16);
+            q13s16 = vsubq_s16(q13s16, q10s16);
+            q11s16 = vsubq_s16(q14s16, q11s16);
+            STORE_IN_OUTPUT(17, 17, 16, q9s16, q8s16)
+            LOAD_FROM_OUTPUT(16, 30, 31, q14s16, q9s16)
+            q8s16  = vsubq_s16(q9s16, q12s16);
+            q10s16 = vaddq_s16(q14s16, q15s16);
+            q14s16 = vsubq_s16(q14s16, q15s16);
+            q12s16 = vaddq_s16(q9s16, q12s16);
+            STORE_IN_OUTPUT(31, 30, 31, q10s16, q12s16)
+            // part of stage 7
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(31, 25, 22, q14s16, q13s16)
+            q13s16 = q11s16;
+            q14s16 = q8s16;
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(22, 24, 23, q14s16, q13s16)
+            // part of stage 4
+            q14s16 = vsubq_s16(q5s16, q0s16);
+            q13s16 = vsubq_s16(q6s16, q2s16);
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q5s16, &q6s16);
+            q14s16 = vsubq_s16(q7s16, q1s16);
+            q13s16 = vsubq_s16(q4s16, q3s16);
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q0s16, &q1s16);
+            // part of stage 6
+            LOAD_FROM_OUTPUT(23, 18, 19, q14s16, q13s16)
+            q8s16 = vaddq_s16(q14s16, q1s16);
+            q9s16 = vaddq_s16(q13s16, q6s16);
+            q13s16 = vsubq_s16(q13s16, q6s16);
+            q1s16 = vsubq_s16(q14s16, q1s16);
+            STORE_IN_OUTPUT(19, 18, 19, q8s16, q9s16)
+            LOAD_FROM_OUTPUT(19, 28, 29, q8s16, q9s16)
+            q14s16 = vsubq_s16(q8s16, q5s16);
+            q10s16 = vaddq_s16(q8s16, q5s16);
+            q11s16 = vaddq_s16(q9s16, q0s16);
+            q0s16 = vsubq_s16(q9s16, q0s16);
+            STORE_IN_OUTPUT(29, 28, 29, q10s16, q11s16)
+            // part of stage 7
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+            STORE_IN_OUTPUT(29, 20, 27, q13s16, q14s16)
+            DO_BUTTERFLY(q0s16, q1s16, cospi_16_64, cospi_16_64,
+                                                         &q1s16, &q0s16);
+            STORE_IN_OUTPUT(27, 21, 26, q1s16, q0s16)
+
+            // -----------------------------------------
+            // BLOCK C: 8-10,11-15
+            // -----------------------------------------
+            // generate 8,9,14,15
+            // part of stage 2
+            LOAD_FROM_TRANSPOSED(3, 2, 30)
+            DO_BUTTERFLY_STD(cospi_30_64, cospi_2_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(30, 18, 14)
+            DO_BUTTERFLY_STD(cospi_14_64, cospi_18_64, &q1s16, &q3s16)
+            // part of stage 3
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 4
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q1s16, &q3s16)
+
+            // generate 10,11,12,13
+            // part of stage 2
+            LOAD_FROM_TRANSPOSED(14, 10, 22)
+            DO_BUTTERFLY_STD(cospi_22_64, cospi_10_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(22, 26, 6)
+            DO_BUTTERFLY_STD(cospi_6_64, cospi_26_64, &q4s16, &q6s16)
+            // part of stage 3
+            q14s16 = vsubq_s16(q4s16, q5s16);
+            q5s16 = vaddq_s16(q4s16, q5s16);
+            q13s16 = vsubq_s16(q6s16, q7s16);
+            q6s16 = vaddq_s16(q6s16, q7s16);
+            // part of stage 4
+            DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q4s16, &q7s16)
+            // part of stage 5
+            q8s16 = vaddq_s16(q0s16, q5s16);
+            q9s16 = vaddq_s16(q1s16, q7s16);
+            q13s16 = vsubq_s16(q1s16, q7s16);
+            q14s16 = vsubq_s16(q3s16, q4s16);
+            q10s16 = vaddq_s16(q3s16, q4s16);
+            q15s16 = vaddq_s16(q2s16, q6s16);
+            STORE_IN_OUTPUT(26, 8, 15, q8s16, q15s16)
+            STORE_IN_OUTPUT(15, 9, 14, q9s16, q10s16)
+            // part of stage 6
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+            STORE_IN_OUTPUT(14, 13, 10, q3s16, q1s16)
+            q13s16 = vsubq_s16(q0s16, q5s16);
+            q14s16 = vsubq_s16(q2s16, q6s16);
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+            STORE_IN_OUTPUT(10, 11, 12, q1s16, q3s16)
+
+            // -----------------------------------------
+            // BLOCK D: 0-3,4-7
+            // -----------------------------------------
+            // generate 4,5,6,7
+            // part of stage 3
+            LOAD_FROM_TRANSPOSED(6, 4, 28)
+            DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q0s16, &q2s16)
+            LOAD_FROM_TRANSPOSED(28, 20, 12)
+            DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+            // part of stage 4
+            q13s16 = vsubq_s16(q0s16, q1s16);
+            q0s16 = vaddq_s16(q0s16, q1s16);
+            q14s16 = vsubq_s16(q2s16, q3s16);
+            q2s16 = vaddq_s16(q2s16, q3s16);
+            // part of stage 5
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+
+            // generate 0,1,2,3
+            // part of stage 4
+            LOAD_FROM_TRANSPOSED(12, 0, 16)
+            DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q5s16, &q7s16)
+            LOAD_FROM_TRANSPOSED(16, 8, 24)
+            DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q14s16, &q6s16)
+            // part of stage 5
+            q4s16 = vaddq_s16(q7s16, q6s16);
+            q7s16 = vsubq_s16(q7s16, q6s16);
+            q6s16 = vsubq_s16(q5s16, q14s16);
+            q5s16 = vaddq_s16(q5s16, q14s16);
+            // part of stage 6
+            q8s16 = vaddq_s16(q4s16, q2s16);
+            q9s16 = vaddq_s16(q5s16, q3s16);
+            q10s16 = vaddq_s16(q6s16, q1s16);
+            q11s16 = vaddq_s16(q7s16, q0s16);
+            q12s16 = vsubq_s16(q7s16, q0s16);
+            q13s16 = vsubq_s16(q6s16, q1s16);
+            q14s16 = vsubq_s16(q5s16, q3s16);
+            q15s16 = vsubq_s16(q4s16, q2s16);
+            // part of stage 7
+            LOAD_FROM_OUTPUT(12, 14, 15, q0s16, q1s16)
+            q2s16 = vaddq_s16(q8s16, q1s16);
+            q3s16 = vaddq_s16(q9s16, q0s16);
+            q4s16 = vsubq_s16(q9s16, q0s16);
+            q5s16 = vsubq_s16(q8s16, q1s16);
+            LOAD_FROM_OUTPUT(15, 16, 17, q0s16, q1s16)
+            q8s16 = vaddq_s16(q4s16, q1s16);
+            q9s16 = vaddq_s16(q5s16, q0s16);
+            q6s16 = vsubq_s16(q5s16, q0s16);
+            q7s16 = vsubq_s16(q4s16, q1s16);
+
+            if (idct32_pass_loop == 0) {
+                idct32_bands_end_1st_pass(out,
+                         q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                         q10s16, q11s16, q12s16, q13s16, q14s16, q15s16);
+            } else {
+                idct32_bands_end_2nd_pass(out, dest, stride,
+                         q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                         q10s16, q11s16, q12s16, q13s16, q14s16, q15s16);
+                dest += 8;
+            }
+        }
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon_asm.asm
new file mode 100644
index 000000000..72e933eee
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct32x32_add_neon_asm.asm
@@ -0,0 +1,1299 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+;TODO(cd): adjust these constant to be able to use vqdmulh for faster
+;          dct_const_round_shift(a * b) within butterfly calculations.
+cospi_1_64  EQU 16364
+cospi_2_64  EQU 16305
+cospi_3_64  EQU 16207
+cospi_4_64  EQU 16069
+cospi_5_64  EQU 15893
+cospi_6_64  EQU 15679
+cospi_7_64  EQU 15426
+cospi_8_64  EQU 15137
+cospi_9_64  EQU 14811
+cospi_10_64 EQU 14449
+cospi_11_64 EQU 14053
+cospi_12_64 EQU 13623
+cospi_13_64 EQU 13160
+cospi_14_64 EQU 12665
+cospi_15_64 EQU 12140
+cospi_16_64 EQU 11585
+cospi_17_64 EQU 11003
+cospi_18_64 EQU 10394
+cospi_19_64 EQU  9760
+cospi_20_64 EQU  9102
+cospi_21_64 EQU  8423
+cospi_22_64 EQU  7723
+cospi_23_64 EQU  7005
+cospi_24_64 EQU  6270
+cospi_25_64 EQU  5520
+cospi_26_64 EQU  4756
+cospi_27_64 EQU  3981
+cospi_28_64 EQU  3196
+cospi_29_64 EQU  2404
+cospi_30_64 EQU  1606
+cospi_31_64 EQU   804
+
+
+    EXPORT  |vp9_idct32x32_1024_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY
+
+    ; --------------------------------------------------------------------------
+    ; Load from transposed_buffer
+    ;   q13 = transposed_buffer[first_offset]
+    ;   q14 = transposed_buffer[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   transposed_buffer must be passed in. use 0 for first use.
+    MACRO
+    LOAD_FROM_TRANSPOSED $prev_offset, $first_offset, $second_offset
+    ; address calculation with proper stride and loading
+    add r0, #($first_offset  - $prev_offset )*8*2
+    vld1.s16        {q14}, [r0]
+    add r0, #($second_offset - $first_offset)*8*2
+    vld1.s16        {q13}, [r0]
+    ; (used) two registers (q14, q13)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Load from output (used as temporary storage)
+    ;   reg1 = output[first_offset]
+    ;   reg2 = output[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    LOAD_FROM_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and loading
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vld1.s16        {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vld1.s16        {$reg2}, [r1]
+    ; (used) two registers ($reg1, $reg2)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Store into output (sometimes as as temporary storage)
+    ;   output[first_offset] = reg1
+    ;   output[second_offset] = reg2
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    STORE_IN_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and storing
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vst1.16 {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vst1.16 {$reg2}, [r1]
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]
+    vst1.16         {d11}, [r9]
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]!
+    vst1.16         {d11}, [r9]!
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]
+    vst1.16         {d4}, [r7]
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]!
+    vst1.16         {d4}, [r7]!
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY $regC, $regD, $regA, $regB, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    ; TODO(cd): have special case to re-use constants when they are similar for
+    ;           consecutive butterflies
+    ; TODO(cd): have special case when both constants are the same, do the
+    ;           additions/subtractions before the multiplies.
+    ; generate the constants
+    ;   generate scalar constants
+    mov             r8,  #$first_constant  & 0xFF00
+    mov             r12, #$second_constant & 0xFF00
+    add             r8,  #$first_constant  & 0x00FF
+    add             r12, #$second_constant & 0x00FF
+    ;   generate vector constants
+    vdup.16         d30, r8
+    vdup.16         d31, r12
+    ; (used) two for inputs (regA-regD), one for constants (q15)
+    ; do some multiplications (ordered for maximum latency hiding)
+    vmull.s16 q8,  $regC, d30
+    vmull.s16 q10, $regA, d31
+    vmull.s16 q9,  $regD, d30
+    vmull.s16 q11, $regB, d31
+    vmull.s16 q12, $regC, d31
+    ; (used) five for intermediate (q8-q12), one for constants (q15)
+    ; do some addition/subtractions (to get back two register)
+    vsub.s32  q8, q8, q10
+    vsub.s32  q9, q9, q11
+    ; do more multiplications (ordered for maximum latency hiding)
+    vmull.s16 q10, $regD, d31
+    vmull.s16 q11, $regA, d30
+    vmull.s16 q15, $regB, d30
+    ; (used) six for intermediate (q8-q12, q15)
+    ; do more addition/subtractions
+    vadd.s32  q11, q12, q11
+    vadd.s32  q10, q10, q15
+    ; (used) four for intermediate (q8-q11)
+    ; dct_const_round_shift
+    vqrshrn.s32 $reg1, q8,  #14
+    vqrshrn.s32 $reg2, q9,  #14
+    vqrshrn.s32 $reg3, q11, #14
+    vqrshrn.s32 $reg4, q10, #14
+    ; (used) two for results, well four d registers
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY_STD $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    DO_BUTTERFLY d28, d29, d26, d27, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    MEND
+    ; --------------------------------------------------------------------------
+
+;void vp9_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int dest_stride);
+;
+;   r0  int16_t *input,
+;   r1  uint8_t *dest,
+;   r2  int dest_stride)
+; loop counters
+;   r4  bands loop counter
+;   r5  pass loop counter
+;   r8  transpose loop counter
+; combine-add pointers
+;   r6  dest + 31 * dest_stride, descending (30, 29, 28, ...)
+;   r7  dest +  0 * dest_stride, ascending  (1, 2, 3, ...)
+;   r9  dest + 15 * dest_stride, descending (14, 13, 12, ...)
+;   r10 dest + 16 * dest_stride, ascending  (17, 18, 19, ...)
+
+|vp9_idct32x32_1024_add_neon| PROC
+    ; This function does one pass of idct32x32 transform.
+    ;
+    ; This is done by transposing the input and then doing a 1d transform on
+    ; columns. In the first pass, the transposed columns are the original
+    ; rows. In the second pass, after the transposition, the colums are the
+    ; original columns.
+    ; The 1d transform is done by looping over bands of eight columns (the
+    ; idct32_bands loop). For each band, the transform input transposition
+    ; is done on demand, one band of four 8x8 matrices at a time. The four
+    ; matrices are transposed by pairs (the idct32_transpose_pair loop).
+    push  {r4-r11}
+    vpush {d8-d15}
+    ; stack operation
+    ; internal buffer used to transpose 8 lines into before transforming them
+    ;   int16_t transpose_buffer[32 * 8];
+    ;   at sp + [4096, 4607]
+    ; results of the first pass (transpose and transform rows)
+    ;   int16_t pass1[32 * 32];
+    ;   at sp + [0, 2047]
+    ; results of the second pass (transpose and transform columns)
+    ;   int16_t pass2[32 * 32];
+    ;   at sp + [2048, 4095]
+    sub sp, sp, #512+2048+2048
+
+    ; r6  = dest + 31 * dest_stride
+    ; r7  = dest +  0 * dest_stride
+    ; r9  = dest + 15 * dest_stride
+    ; r10 = dest + 16 * dest_stride
+    rsb r6,  r2, r2, lsl #5
+    rsb r9,  r2, r2, lsl #4
+    add r10, r1, r2, lsl #4
+    mov r7, r1
+    add r6, r6, r1
+    add r9, r9, r1
+    ; r11 = -dest_stride
+    neg r11, r2
+    ; r3 = input
+    mov r3, r0
+    ; parameters for first pass
+      ; r0 = transpose_buffer[32 * 8]
+    add r0, sp, #4096
+      ; r1 = pass1[32 * 32]
+    mov r1, sp
+
+    mov r5, #0          ; initialize pass loop counter
+idct32_pass_loop
+    mov r4, #4          ; initialize bands loop counter
+idct32_bands_loop
+    mov r8, #2          ; initialize transpose loop counter
+idct32_transpose_pair_loop
+    ; Load two horizontally consecutive 8x8 16bit data matrices. The first one
+    ; into q0-q7 and the second one into q8-q15. There is a stride of 64,
+    ; adjusted to 32 because of the two post-increments.
+    vld1.s16        {q8},  [r3]!
+    vld1.s16        {q0},  [r3]!
+    add r3, #32
+    vld1.s16        {q9},  [r3]!
+    vld1.s16        {q1},  [r3]!
+    add r3, #32
+    vld1.s16        {q10}, [r3]!
+    vld1.s16        {q2},  [r3]!
+    add r3, #32
+    vld1.s16        {q11}, [r3]!
+    vld1.s16        {q3},  [r3]!
+    add r3, #32
+    vld1.s16        {q12}, [r3]!
+    vld1.s16        {q4},  [r3]!
+    add r3, #32
+    vld1.s16        {q13}, [r3]!
+    vld1.s16        {q5},  [r3]!
+    add r3, #32
+    vld1.s16        {q14}, [r3]!
+    vld1.s16        {q6},  [r3]!
+    add r3, #32
+    vld1.s16        {q15}, [r3]!
+    vld1.s16        {q7},  [r3]!
+
+    ; Transpose the two 8x8 16bit data matrices.
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vswp            d1,  d8
+    vswp            d7,  d14
+    vswp            d5,  d12
+    vswp            d3,  d10
+    vtrn.32         q8,  q10
+    vtrn.32         q9,  q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.32         q0,  q2
+    vtrn.32         q1,  q3
+    vtrn.32         q4,  q6
+    vtrn.32         q5,  q7
+    vtrn.16         q8,  q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    vtrn.16         q0,  q1
+    vtrn.16         q2,  q3
+    vtrn.16         q4,  q5
+    vtrn.16         q6,  q7
+
+    ; Store both matrices after each other. There is a stride of 32, which
+    ; adjusts to nothing because of the post-increments.
+    vst1.16        {q8},  [r0]!
+    vst1.16        {q9},  [r0]!
+    vst1.16        {q10}, [r0]!
+    vst1.16        {q11}, [r0]!
+    vst1.16        {q12}, [r0]!
+    vst1.16        {q13}, [r0]!
+    vst1.16        {q14}, [r0]!
+    vst1.16        {q15}, [r0]!
+    vst1.16        {q0},  [r0]!
+    vst1.16        {q1},  [r0]!
+    vst1.16        {q2},  [r0]!
+    vst1.16        {q3},  [r0]!
+    vst1.16        {q4},  [r0]!
+    vst1.16        {q5},  [r0]!
+    vst1.16        {q6},  [r0]!
+    vst1.16        {q7},  [r0]!
+
+    ; increment pointers by adjusted stride (not necessary for r0/out)
+    ;   go back by 7*32 for the seven lines moved fully by read and add
+    ;   go back by 32 for the eigth line only read
+    ;   advance by 16*2 to go the next pair
+    sub r3,  r3,  #7*32*2 + 32 - 16*2
+    ; transpose pair loop processing
+    subs r8, r8, #1
+    bne idct32_transpose_pair_loop
+
+    ; restore r0/input to its original value
+    sub r0, r0, #32*8*2
+
+    ; Instead of doing the transforms stage by stage, it is done by loading
+    ; some input values and doing as many stages as possible to minimize the
+    ; storing/loading of intermediate results. To fit within registers, the
+    ; final coefficients are cut into four blocks:
+    ; BLOCK A: 16-19,28-31
+    ; BLOCK B: 20-23,24-27
+    ; BLOCK C: 8-10,11-15
+    ; BLOCK D: 0-3,4-7
+    ; Blocks A and C are straight calculation through the various stages. In
+    ; block B, further calculations are performed using the results from
+    ; block A. In block D, further calculations are performed using the results
+    ; from block C and then the final calculations are done using results from
+    ; block A and B which have been combined at the end of block B.
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK A: 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; generate 16,17,30,31
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[1 * 32] * cospi_31_64 - input[31 * 32] *  cospi_1_64;
+    ;temp2 = input[1 * 32] *  cospi_1_64 + input[31 * 32] * cospi_31_64;
+    ;step1b[16][i] = dct_const_round_shift(temp1);
+    ;step1b[31][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 0, 1, 31
+    DO_BUTTERFLY_STD cospi_31_64, cospi_1_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[17 * 32] * cospi_15_64 - input[15 * 32] * cospi_17_64;
+    ;temp2 = input[17 * 32] * cospi_17_64 + input[15 * 32] * cospi_15_64;
+    ;step1b[17][i] = dct_const_round_shift(temp1);
+    ;step1b[30][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 31, 17, 15
+    DO_BUTTERFLY_STD cospi_15_64, cospi_17_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[16] =  step1b[16][i] + step1b[17][i];
+    ;step2[17] =  step1b[16][i] - step1b[17][i];
+    ;step2[30] = -step1b[30][i] + step1b[31][i];
+    ;step2[31] =  step1b[30][i] + step1b[31][i];
+    vadd.s16  q4, q0, q1
+    vsub.s16  q13, q0, q1
+    vadd.s16  q6, q2, q3
+    vsub.s16  q14, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[30][i] * cospi_28_64 - step1b[17][i] * cospi_4_64;
+    ;temp2 = step1b[30][i] * cospi_4_64  - step1b[17][i] * cospi_28_64;
+    ;step3[17] = dct_const_round_shift(temp1);
+    ;step3[30] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; generate 18,19,28,29
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[9 * 32] * cospi_23_64 - input[23 * 32] * cospi_9_64;
+    ;temp2 = input[9 * 32] *  cospi_9_64 + input[23 * 32] * cospi_23_64;
+    ;step1b[18][i] = dct_const_round_shift(temp1);
+    ;step1b[29][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 15, 9, 23
+    DO_BUTTERFLY_STD cospi_23_64, cospi_9_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[25 * 32] *  cospi_7_64 - input[7 * 32] * cospi_25_64;
+    ;temp2 = input[25 * 32] * cospi_25_64 + input[7 * 32] * cospi_7_64;
+    ;step1b[19][i] = dct_const_round_shift(temp1);
+    ;step1b[28][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 23, 25, 7
+    DO_BUTTERFLY_STD cospi_7_64, cospi_25_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[18] = -step1b[18][i] + step1b[19][i];
+    ;step2[19] =  step1b[18][i] + step1b[19][i];
+    ;step2[28] =  step1b[28][i] + step1b[29][i];
+    ;step2[29] =  step1b[28][i] - step1b[29][i];
+    vsub.s16  q13, q3, q2
+    vadd.s16  q3,  q3, q2
+    vsub.s16  q14, q1, q0
+    vadd.s16  q2,  q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[18][i] * (-cospi_4_64)  - step1b[29][i] * (-cospi_28_64);
+    ;temp2 = step1b[18][i] * (-cospi_28_64) + step1b[29][i] * (-cospi_4_64);
+    ;step3[29] = dct_const_round_shift(temp1);
+    ;step3[18] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_4_64), (-cospi_28_64), d2, d3, d0, d1
+    ; --------------------------------------------------------------------------
+    ; combine 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[16] = step1b[16][i] + step1b[19][i];
+    ;step1[17] = step1b[17][i] + step1b[18][i];
+    ;step1[18] = step1b[17][i] - step1b[18][i];
+    ;step1[29] = step1b[30][i] - step1b[29][i];
+    ;step1[30] = step1b[30][i] + step1b[29][i];
+    ;step1[31] = step1b[31][i] + step1b[28][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q0
+    vadd.s16  q10, q7, q1
+    vadd.s16  q15, q6, q3
+    vsub.s16  q13, q5, q0
+    vsub.s16  q14, q7, q1
+    STORE_IN_OUTPUT 0,  16, 31, q8,  q15
+    STORE_IN_OUTPUT 31, 17, 30, q9,  q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[29][i] * cospi_24_64 - step1b[18][i] * cospi_8_64;
+    ;temp2 = step1b[29][i] * cospi_8_64  + step1b[18][i] * cospi_24_64;
+    ;step2[18] = dct_const_round_shift(temp1);
+    ;step2[29] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d0, d1, d2, d3
+    STORE_IN_OUTPUT 30, 29, 18, q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[19] = step1b[16][i] - step1b[19][i];
+    ;step1[28] = step1b[31][i] - step1b[28][i];
+    vsub.s16  q13, q4, q2
+    vsub.s16  q14, q6, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[28][i] * cospi_24_64 - step1b[19][i] * cospi_8_64;
+    ;temp2 = step1b[28][i] * cospi_8_64  + step1b[19][i] * cospi_24_64;
+    ;step2[19] = dct_const_round_shift(temp1);
+    ;step2[28] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d8, d9, d12, d13
+    STORE_IN_OUTPUT 18, 19, 28, q4, q6
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK B: 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; generate 20,21,26,27
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[5 * 32] * cospi_27_64 - input[27 * 32] * cospi_5_64;
+    ;temp2 = input[5 * 32] *  cospi_5_64 + input[27 * 32] * cospi_27_64;
+    ;step1b[20][i] = dct_const_round_shift(temp1);
+    ;step1b[27][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 7, 5, 27
+    DO_BUTTERFLY_STD cospi_27_64, cospi_5_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[21 * 32] * cospi_11_64 - input[11 * 32] * cospi_21_64;
+    ;temp2 = input[21 * 32] * cospi_21_64 + input[11 * 32] * cospi_11_64;
+    ;step1b[21][i] = dct_const_round_shift(temp1);
+    ;step1b[26][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 27, 21, 11
+    DO_BUTTERFLY_STD cospi_11_64, cospi_21_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[20] =  step1b[20][i] + step1b[21][i];
+    ;step2[21] =  step1b[20][i] - step1b[21][i];
+    ;step2[26] = -step1b[26][i] + step1b[27][i];
+    ;step2[27] =  step1b[26][i] + step1b[27][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[26][i] * cospi_12_64 - step1b[21][i] * cospi_20_64;
+    ;temp2 = step1b[26][i] * cospi_20_64 + step1b[21][i] * cospi_12_64;
+    ;step3[21] = dct_const_round_shift(temp1);
+    ;step3[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 22,23,24,25
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[13 * 32] * cospi_19_64 - input[19 * 32] * cospi_13_64;
+    ;temp2 = input[13 * 32] * cospi_13_64 + input[19 * 32] * cospi_19_64;
+    ;step1b[22][i] = dct_const_round_shift(temp1);
+    ;step1b[25][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 11, 13, 19
+    DO_BUTTERFLY_STD cospi_19_64, cospi_13_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[29 * 32] *  cospi_3_64 - input[3 * 32] * cospi_29_64;
+    ;temp2 = input[29 * 32] * cospi_29_64 + input[3 * 32] * cospi_3_64;
+    ;step1b[23][i] = dct_const_round_shift(temp1);
+    ;step1b[24][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 19, 29, 3
+    DO_BUTTERFLY_STD cospi_3_64, cospi_29_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[22] = -step1b[22][i] + step1b[23][i];
+    ;step2[23] =  step1b[22][i] + step1b[23][i];
+    ;step2[24] =  step1b[24][i] + step1b[25][i];
+    ;step2[25] =  step1b[24][i] - step1b[25][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[22][i] * (-cospi_20_64) - step1b[25][i] * (-cospi_12_64);
+    ;temp2 = step1b[22][i] * (-cospi_12_64) + step1b[25][i] * (-cospi_20_64);
+    ;step3[25] = dct_const_round_shift(temp1);
+    ;step3[22] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_20_64), (-cospi_12_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[22] = step1b[22][i] + step1b[21][i];
+    ;step1[23] = step1b[23][i] + step1b[20][i];
+    vadd.s16  q10, q7, q1
+    vadd.s16  q11, q5, q0
+    ;step1[24] = step1b[24][i] + step1b[27][i];
+    ;step1[25] = step1b[25][i] + step1b[26][i];
+    vadd.s16  q12, q6, q2
+    vadd.s16  q15, q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[16] = step1b[16][i] + step1b[23][i];
+    ;step3[17] = step1b[17][i] + step1b[22][i];
+    ;step3[22] = step1b[17][i] - step1b[22][i];
+    ;step3[23] = step1b[16][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 28, 16, 17, q14, q13
+    vadd.s16  q8,  q14, q11
+    vadd.s16  q9,  q13, q10
+    vsub.s16  q13, q13, q10
+    vsub.s16  q11, q14, q11
+    STORE_IN_OUTPUT 17, 17, 16, q9, q8
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[24] = step1b[31][i] - step1b[24][i];
+    ;step3[25] = step1b[30][i] - step1b[25][i];
+    ;step3[30] = step1b[30][i] + step1b[25][i];
+    ;step3[31] = step1b[31][i] + step1b[24][i];
+    LOAD_FROM_OUTPUT 16, 30, 31, q14, q9
+    vsub.s16  q8,  q9,  q12
+    vadd.s16  q10, q14, q15
+    vsub.s16  q14, q14, q15
+    vadd.s16  q12, q9,  q12
+    STORE_IN_OUTPUT 31, 30, 31, q10, q12
+    ; --------------------------------------------------------------------------
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpush {q8}  ; [24]
+    vpush {q11} ; [23]
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[25][i] - step1b[22][i]) * cospi_16_64;
+    ;temp2 = (step1b[25][i] + step1b[22][i]) * cospi_16_64;
+    ;step1[22] = dct_const_round_shift(temp1);
+    ;step1[25] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 31, 25, 22, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[24][i] - step1b[23][i]) * cospi_16_64;
+    ;temp2 = (step1b[24][i] + step1b[23][i]) * cospi_16_64;
+    ;step1[23] = dct_const_round_shift(temp1);
+    ;step1[24] = dct_const_round_shift(temp2);
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpop  {q13} ; [23]
+    vpop  {q14} ; [24]
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 22, 24, 23, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[20] = step1b[23][i] - step1b[20][i];
+    ;step1[27] = step1b[24][i] - step1b[27][i];
+    vsub.s16  q14, q5, q0
+    vsub.s16  q13, q6, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[20][i] * (-cospi_8_64)  - step1b[27][i] * (-cospi_24_64);
+    ;temp2 = step1b[20][i] * (-cospi_24_64) + step1b[27][i] * (-cospi_8_64);
+    ;step2[27] = dct_const_round_shift(temp1);
+    ;step2[20] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d10, d11, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[21] = step1b[22][i] - step1b[21][i];
+    ;step1[26] = step1b[25][i] - step1b[26][i];
+    vsub.s16  q14,  q7, q1
+    vsub.s16  q13,  q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[21][i] * (-cospi_8_64)  - step1b[26][i] * (-cospi_24_64);
+    ;temp2 = step1b[21][i] * (-cospi_24_64) + step1b[26][i] * (-cospi_8_64);
+    ;step2[26] = dct_const_round_shift(temp1);
+    ;step2[21] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d0, d1, d2, d3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[18] = step1b[18][i] + step1b[21][i];
+    ;step3[19] = step1b[19][i] + step1b[20][i];
+    ;step3[20] = step1b[19][i] - step1b[20][i];
+    ;step3[21] = step1b[18][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 23, 18, 19, q14, q13
+    vadd.s16  q8,  q14, q1
+    vadd.s16  q9,  q13, q6
+    vsub.s16  q13, q13, q6
+    vsub.s16  q1,  q14, q1
+    STORE_IN_OUTPUT 19, 18, 19, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[27] = step1b[28][i] - step1b[27][i];
+    ;step3[28] = step1b[28][i] + step1b[27][i];
+    ;step3[29] = step1b[29][i] + step1b[26][i];
+    ;step3[26] = step1b[29][i] - step1b[26][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q8, q9
+    vsub.s16  q14, q8, q5
+    vadd.s16  q10, q8, q5
+    vadd.s16  q11, q9, q0
+    vsub.s16  q0, q9, q0
+    STORE_IN_OUTPUT 29, 28, 29, q10, q11
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[27][i] - step1b[20][i]) * cospi_16_64;
+    ;temp2 = (step1b[27][i] + step1b[20][i]) * cospi_16_64;
+    ;step1[20] = dct_const_round_shift(temp1);
+    ;step1[27] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 29, 20, 27, q13, q14
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[26][i] - step1b[21][i]) * cospi_16_64;
+    ;temp2 = (step1b[26][i] + step1b[21][i]) * cospi_16_64;
+    ;step1[21] = dct_const_round_shift(temp1);
+    ;step1[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY d0, d1, d2, d3, cospi_16_64, cospi_16_64, d2, d3, d0, d1
+    STORE_IN_OUTPUT 27, 21, 26, q1, q0
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK C: 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; generate 8,9,14,15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[2 * 32] * cospi_30_64 - input[30 * 32] * cospi_2_64;
+    ;temp2 = input[2 * 32] * cospi_2_64 + input[30 * 32] * cospi_30_64;
+    ;step2[8] = dct_const_round_shift(temp1);
+    ;step2[15] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 3, 2, 30
+    DO_BUTTERFLY_STD cospi_30_64, cospi_2_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[18 * 32] * cospi_14_64 - input[14 * 32] * cospi_18_64;
+    ;temp2 = input[18 * 32] * cospi_18_64 + input[14 * 32] * cospi_14_64;
+    ;step2[9] = dct_const_round_shift(temp1);
+    ;step2[14] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 30, 18, 14
+    DO_BUTTERFLY_STD cospi_14_64, cospi_18_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[8] = step1b[8][i] + step1b[9][i];
+    ;step3[9] = step1b[8][i] - step1b[9][i];
+    ;step3[14] = step1b[15][i] - step1b[14][i];
+    ;step3[15] = step1b[15][i] + step1b[14][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[14][i] * cospi_24_64 - step1b[9][i] * cospi_8_64;
+    ;temp2 = step1b[14][i] * cospi_8_64  + step1b[9][i] * cospi_24_64;
+    ;step1[9]  = dct_const_round_shift(temp1);
+    ;step1[14] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 10,11,12,13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[10 * 32] * cospi_22_64 - input[22 * 32] * cospi_10_64;
+    ;temp2 = input[10 * 32] * cospi_10_64 + input[22 * 32] * cospi_22_64;
+    ;step2[10] = dct_const_round_shift(temp1);
+    ;step2[13] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 14, 10, 22
+    DO_BUTTERFLY_STD cospi_22_64, cospi_10_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[26 * 32] * cospi_6_64 - input[6 * 32] * cospi_26_64;
+    ;temp2 = input[26 * 32] * cospi_26_64 + input[6 * 32] * cospi_6_64;
+    ;step2[11] = dct_const_round_shift(temp1);
+    ;step2[12] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 22, 26, 6
+    DO_BUTTERFLY_STD cospi_6_64, cospi_26_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[10] = step1b[11][i] - step1b[10][i];
+    ;step3[11] = step1b[11][i] + step1b[10][i];
+    ;step3[12] = step1b[12][i] + step1b[13][i];
+    ;step3[13] = step1b[12][i] - step1b[13][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[10][i] * (-cospi_8_64)  - step1b[13][i] * (-cospi_24_64);
+    ;temp2 = step1b[10][i] * (-cospi_24_64) + step1b[13][i] * (-cospi_8_64);
+    ;step1[13] = dct_const_round_shift(temp1);
+    ;step1[10] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[8]  = step1b[8][i] + step1b[11][i];
+    ;step2[9]  = step1b[9][i] + step1b[10][i];
+    ;step2[10] = step1b[9][i] - step1b[10][i];
+    vadd.s16  q8,  q0, q5
+    vadd.s16  q9,  q1, q7
+    vsub.s16  q13, q1, q7
+    ;step2[13] = step1b[14][i] - step1b[13][i];
+    ;step2[14] = step1b[14][i] + step1b[13][i];
+    ;step2[15] = step1b[15][i] + step1b[12][i];
+    vsub.s16  q14, q3, q4
+    vadd.s16  q10, q3, q4
+    vadd.s16  q15, q2, q6
+    STORE_IN_OUTPUT 26, 8, 15, q8, q15
+    STORE_IN_OUTPUT 15, 9, 14, q9, q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[13][i] - step1b[10][i]) * cospi_16_64;
+    ;temp2 = (step1b[13][i] + step1b[10][i]) * cospi_16_64;
+    ;step3[10] = dct_const_round_shift(temp1);
+    ;step3[13] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 14, 13, 10, q3, q1
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[11] = step1b[8][i] - step1b[11][i];
+    ;step2[12] = step1b[15][i] - step1b[12][i];
+    vsub.s16  q13, q0, q5
+    vsub.s16  q14,  q2, q6
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[12][i] - step1b[11][i]) * cospi_16_64;
+    ;temp2 = (step1b[12][i] + step1b[11][i]) * cospi_16_64;
+    ;step3[11] = dct_const_round_shift(temp1);
+    ;step3[12] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 10, 11, 12, q1, q3
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK D: 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; generate 4,5,6,7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[4 * 32] * cospi_28_64 - input[28 * 32] * cospi_4_64;
+    ;temp2 = input[4 * 32] * cospi_4_64 + input[28 * 32] * cospi_28_64;
+    ;step3[4] = dct_const_round_shift(temp1);
+    ;step3[7] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 6, 4, 28
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[20 * 32] * cospi_12_64 - input[12 * 32] * cospi_20_64;
+    ;temp2 = input[20 * 32] * cospi_20_64 + input[12 * 32] * cospi_12_64;
+    ;step3[5] = dct_const_round_shift(temp1);
+    ;step3[6] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 28, 20, 12
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[4] = step1b[4][i] + step1b[5][i];
+    ;step1[5] = step1b[4][i] - step1b[5][i];
+    ;step1[6] = step1b[7][i] - step1b[6][i];
+    ;step1[7] = step1b[7][i] + step1b[6][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = (step1b[6][i] - step1b[5][i]) * cospi_16_64;
+    ;temp2 = (step1b[5][i] + step1b[6][i]) * cospi_16_64;
+    ;step2[5] = dct_const_round_shift(temp1);
+    ;step2[6] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 0,1,2,3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = (input[0 * 32] - input[16 * 32]) * cospi_16_64;
+    ;temp2 = (input[0 * 32] + input[16 * 32]) * cospi_16_64;
+    ;step1[1] = dct_const_round_shift(temp1);
+    ;step1[0] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 12, 0, 16
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = input[8 * 32] * cospi_24_64 - input[24 * 32] * cospi_8_64;
+    ;temp2 = input[8 * 32] * cospi_8_64 + input[24 * 32] * cospi_24_64;
+    ;step1[2] = dct_const_round_shift(temp1);
+    ;step1[3] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 16, 8, 24
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d28, d29, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[0] = step1b[0][i] + step1b[3][i];
+    ;step2[1] = step1b[1][i] + step1b[2][i];
+    ;step2[2] = step1b[1][i] - step1b[2][i];
+    ;step2[3] = step1b[0][i] - step1b[3][i];
+    vadd.s16  q4, q7, q6
+    vsub.s16  q7, q7, q6
+    vsub.s16  q6, q5, q14
+    vadd.s16  q5, q5, q14
+    ; --------------------------------------------------------------------------
+    ; combine 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[0] = step1b[0][i] + step1b[7][i];
+    ;step3[1] = step1b[1][i] + step1b[6][i];
+    ;step3[2] = step1b[2][i] + step1b[5][i];
+    ;step3[3] = step1b[3][i] + step1b[4][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q3
+    vadd.s16  q10, q6, q1
+    vadd.s16  q11, q7, q0
+    ;step3[4] = step1b[3][i] - step1b[4][i];
+    ;step3[5] = step1b[2][i] - step1b[5][i];
+    ;step3[6] = step1b[1][i] - step1b[6][i];
+    ;step3[7] = step1b[0][i] - step1b[7][i];
+    vsub.s16  q12, q7, q0
+    vsub.s16  q13, q6, q1
+    vsub.s16  q14, q5, q3
+    vsub.s16  q15, q4, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[0] = step1b[0][i] + step1b[15][i];
+    ;step1[1] = step1b[1][i] + step1b[14][i];
+    ;step1[14] = step1b[1][i] - step1b[14][i];
+    ;step1[15] = step1b[0][i] - step1b[15][i];
+    LOAD_FROM_OUTPUT 12, 14, 15, q0, q1
+    vadd.s16  q2, q8, q1
+    vadd.s16  q3, q9, q0
+    vsub.s16  q4, q9, q0
+    vsub.s16  q5, q8, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[14 * 32] = step1b[14][i] + step1b[17][i];
+    ;output[15 * 32] = step1b[15][i] + step1b[16][i];
+    ;output[16 * 32] = step1b[15][i] - step1b[16][i];
+    ;output[17 * 32] = step1b[14][i] - step1b[17][i];
+    LOAD_FROM_OUTPUT 15, 16, 17, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+
+    cmp r5, #0
+    bgt idct32_bands_end_2nd_pass
+
+idct32_bands_end_1st_pass
+    STORE_IN_OUTPUT 17, 16, 17, q6, q7
+    STORE_IN_OUTPUT 17, 14, 15, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 15, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 31, 30, 31, q6, q7
+    STORE_IN_OUTPUT 31,  0,  1, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 1, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 19, 18, 19, q6, q7
+    STORE_IN_OUTPUT 19, 12, 13, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 13, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 29, 28, 29, q6, q7
+    STORE_IN_OUTPUT 29,  2,  3, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 3, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 21, 20, 21, q6, q7
+    STORE_IN_OUTPUT 21, 10, 11, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 11, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 27, 26, 27, q6, q7
+    STORE_IN_OUTPUT 27,  4,  5, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 5, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 23, 22, 23, q6, q7
+    STORE_IN_OUTPUT 23, 8, 9, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 9, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 25, 24, 25, q6, q7
+    STORE_IN_OUTPUT 25,  6,  7, q4, q5
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (STORE_IN_OUTPUT 24,  6,  7) => 7*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #7*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; parameters for second pass
+    ; the input of pass2 is the result of pass1. we have to remove the offset
+    ;   of 32 columns induced by the above idct32_bands_loop
+    sub r3, r1, #32*2
+      ; r1 = pass2[32 * 32]
+    add r1, sp, #2048
+
+    ; pass loop processing
+    add r5, r5, #1
+    b idct32_pass_loop
+
+idct32_bands_end_2nd_pass
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 17, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 31, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 29, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 21, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 27, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 23, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; restore pointers to their initial indices for next band pass by
+    ;     removing/adding dest_stride * 8. The actual increment by eight
+    ;     is taken care of within the _LAST macros.
+    add r6,  r6,  r2, lsl #3
+    add r9,  r9,  r2, lsl #3
+    sub r7,  r7,  r2, lsl #3
+    sub r10, r10, r2, lsl #3
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (LOAD_FROM_OUTPUT 23, 24, 25) => 25*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #25*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; stack operation
+    add sp, sp, #512+2048+2048
+    vpop {d8-d15}
+    pop  {r4-r11}
+    bx              lr
+    ENDP  ; |vp9_idct32x32_1024_add_neon|
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.c
new file mode 100644
index 000000000..f0457358e
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon.c
@@ -0,0 +1,50 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_idct.h"
+
+void vp9_idct4x4_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d6u8;
+    uint32x2_t d2u32 = vdup_n_u32(0);
+    uint16x8_t q8u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 4);
+
+    q0s16 = vdupq_n_s16(a1);
+
+    // dc_only_idct_add
+    d1 = d2 = dest;
+    for (i = 0; i < 2; i++) {
+        d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
+        d1 += dest_stride;
+        d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
+        d1 += dest_stride;
+
+        q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16),
+                         vreinterpret_u8_u32(d2u32));
+        d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+
+        vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
+        d2 += dest_stride;
+        vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon_asm.asm
new file mode 100644
index 000000000..0d4a721c4
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_1_add_neon_asm.asm
@@ -0,0 +1,68 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vp9_idct4x4_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_idct4x4_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct4x4_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 4)
+    add              r0, r0, #8                ; + (1 <<((4) - 1))
+    asr              r0, r0, #4                ; >> 4
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    vld1.32          {d2[0]}, [r1], r2
+    vld1.32          {d2[1]}, [r1], r2
+    vld1.32          {d4[0]}, [r1], r2
+    vld1.32          {d4[1]}, [r1]
+
+    vaddw.u8         q8, q0, d2                ; dest[x] + a1
+    vaddw.u8         q9, q0, d4
+
+    vqmovun.s16      d6, q8                    ; clip_pixel
+    vqmovun.s16      d7, q9
+
+    vst1.32          {d6[0]}, [r12], r2
+    vst1.32          {d6[1]}, [r12], r2
+    vst1.32          {d7[0]}, [r12], r2
+    vst1.32          {d7[1]}, [r12]
+
+    bx               lr
+    ENDP             ; |vp9_idct4x4_1_add_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.c
new file mode 100644
index 000000000..dc91e0f30
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon.c
@@ -0,0 +1,151 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+void vp9_idct4x4_16_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d26u8, d27u8;
+    uint32x2_t d26u32, d27u32;
+    uint16x8_t q8u16, q9u16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
+    int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
+    int16x8_t q8s16, q9s16, q13s16, q14s16;
+    int32x4_t q1s32, q13s32, q14s32, q15s32;
+    int16x4x2_t d0x2s16, d1x2s16;
+    int32x4x2_t q0x2s32;
+    uint8_t *d;
+    int16_t cospi_8_64 = 15137;
+    int16_t cospi_16_64 = 11585;
+    int16_t cospi_24_64 = 6270;
+
+    d26u32 = d27u32 = vdup_n_u32(0);
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+
+    d0x2s16 = vtrn_s16(d16s16, d17s16);
+    d1x2s16 = vtrn_s16(d18s16, d19s16);
+    q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+    q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+    d20s16 = vdup_n_s16(cospi_8_64);
+    d21s16 = vdup_n_s16(cospi_16_64);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                        vreinterpretq_s32_s16(q9s16));
+    d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+    d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+    d22s16 = vdup_n_s16(cospi_24_64);
+
+    // stage 1
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, d22s16);
+    q1s32  = vmull_s16(d17s16, d20s16);
+    q13s32 = vmull_s16(d23s16, d21s16);
+    q14s32 = vmull_s16(d24s16, d21s16);
+
+    q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+    q1s32  = vmlal_s16(q1s32,  d19s16, d22s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q1s32,  14);
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+
+    // stage 2
+    q8s16 = vaddq_s16(q13s16, q14s16);
+    q9s16 = vsubq_s16(q13s16, q14s16);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_high_s16(q9s16);  // vswp d18 d19
+    d19s16 = vget_low_s16(q9s16);
+
+    d0x2s16 = vtrn_s16(d16s16, d17s16);
+    d1x2s16 = vtrn_s16(d18s16, d19s16);
+    q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+    q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(q8s16),
+                        vreinterpretq_s32_s16(q9s16));
+    d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+    d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+    d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+    // do the transform on columns
+    // stage 1
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, d22s16);
+    q1s32  = vmull_s16(d17s16, d20s16);
+    q13s32 = vmull_s16(d23s16, d21s16);
+    q14s32 = vmull_s16(d24s16, d21s16);
+
+    q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+    q1s32  = vmlal_s16(q1s32,  d19s16, d22s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q1s32,  14);
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+
+    // stage 2
+    q8s16 = vaddq_s16(q13s16, q14s16);
+    q9s16 = vsubq_s16(q13s16, q14s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 4);
+    q9s16 = vrshrq_n_s16(q9s16, 4);
+
+    d = dest;
+    d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
+    d += dest_stride;
+    d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
+    d += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
+    d += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u32(d26u32));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u32(d27u32));
+
+    d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+    d = dest;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
+    d += dest_stride;
+    vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon_asm.asm
new file mode 100644
index 000000000..00283fc8d
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct4x4_add_neon_asm.asm
@@ -0,0 +1,190 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_idct4x4_16_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY ; name this block of code
+;void vp9_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct4x4_16_add_neon| PROC
+
+    ; The 2D transform is done with two passes which are actually pretty
+    ; similar. We first transform the rows. This is done by transposing
+    ; the inputs, doing an SIMD column transform (the columns are the
+    ; transposed rows) and then transpose the results (so that it goes back
+    ; in normal/row positions). Then, we transform the columns by doing
+    ; another SIMD column transform.
+    ; So, two passes of a transpose followed by a column transform.
+
+    ; load the inputs into q8-q9, d16-d19
+    vld1.s16        {q8,q9}, [r0]!
+
+    ; generate scalar constants
+    ; cospi_8_64 = 15137 = 0x3b21
+    mov             r0, #0x3b00
+    add             r0, #0x21
+    ; cospi_16_64 = 11585 = 0x2d41
+    mov             r3, #0x2d00
+    add             r3, #0x41
+    ; cospi_24_64 = 6270 = 0x 187e
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; transpose the input data
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+
+    ; generate constant vectors
+    vdup.16         d20, r0         ; replicate cospi_8_64
+    vdup.16         d21, r3         ; replicate cospi_16_64
+
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    vdup.16         d22, r12        ; replicate cospi_24_64
+
+    ; do the transform on transposed rows
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+    vswp     d18, d19
+
+    ; transpose the results
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    ; do the transform on columns
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+
+    ; The results are in two registers, one of them being swapped. This will
+    ; be taken care of by loading the 'dest' value in a swapped fashion and
+    ; also storing them in the same swapped fashion.
+    ; temp_out[0, 1] = d16, d17 = q8
+    ; temp_out[2, 3] = d19, d18 = q9 swapped
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4)
+    vrshr.s16 q8, q8, #4
+    vrshr.s16 q9, q9, #4
+
+    vld1.32 {d26[0]}, [r1], r2
+    vld1.32 {d26[1]}, [r1], r2
+    vld1.32 {d27[1]}, [r1], r2
+    vld1.32 {d27[0]}, [r1]  ; no post-increment
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]
+    vaddw.u8 q8, q8, d26
+    vaddw.u8 q9, q9, d27
+
+    ; clip_pixel
+    vqmovun.s16 d26, q8
+    vqmovun.s16 d27, q9
+
+    ; do the stores in reverse order with negative post-increment, by changing
+    ; the sign of the stride
+    rsb r2, r2, #0
+    vst1.32 {d27[0]}, [r1], r2
+    vst1.32 {d27[1]}, [r1], r2
+    vst1.32 {d26[1]}, [r1], r2
+    vst1.32 {d26[0]}, [r1]  ; no post-increment
+    bx              lr
+    ENDP  ; |vp9_idct4x4_16_add_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.c
new file mode 100644
index 000000000..5369697c7
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon.c
@@ -0,0 +1,64 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_idct.h"
+
+void vp9_idct8x8_1_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8x8_t d2u8, d3u8, d30u8, d31u8;
+    uint64x1_t d2u64, d3u64, d4u64, d5u64;
+    uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+    int16x8_t q0s16;
+    uint8_t *d1, *d2;
+    int16_t i, a1, cospi_16_64 = 11585;
+    int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+    out = dct_const_round_shift(out * cospi_16_64);
+    a1 = ROUND_POWER_OF_TWO(out, 5);
+
+    q0s16 = vdupq_n_s16(a1);
+    q0u16 = vreinterpretq_u16_s16(q0s16);
+
+    d1 = d2 = dest;
+    for (i = 0; i < 2; i++) {
+        d2u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d3u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d4u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+        d5u64 = vld1_u64((const uint64_t *)d1);
+        d1 += dest_stride;
+
+        q9u16  = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+        q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+        q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+        q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+        d2u8  = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        d3u8  = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+        d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+        d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8));
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon_asm.asm
new file mode 100644
index 000000000..421d202d4
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_1_add_neon_asm.asm
@@ -0,0 +1,88 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+    EXPORT  |vp9_idct8x8_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_idct8x8_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct8x8_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 5)
+    add              r0, r0, #16               ; + (1 <<((5) - 1))
+    asr              r0, r0, #5                ; >> 5
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    ; load destination data
+    vld1.64          {d2}, [r1], r2
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r2
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r2
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r2
+    vld1.64          {d17}, [r1]
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |vp9_idct8x8_1_add_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.c
new file mode 100644
index 000000000..2b3c1ce60
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon.c
@@ -0,0 +1,547 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+static int16_t cospi_4_64 = 16069;
+static int16_t cospi_8_64 = 15137;
+static int16_t cospi_12_64 = 13623;
+static int16_t cospi_16_64 = 11585;
+static int16_t cospi_20_64 = 9102;
+static int16_t cospi_24_64 = 6270;
+static int16_t cospi_28_64 = 3196;
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+static INLINE void IDCT8x8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d26s16, d2s16);
+    q6s32 = vmull_s16(d27s16, d2s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+    q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+    d8s16 = vqrshrn_n_s32(q2s32, 14);
+    d9s16 = vqrshrn_n_s32(q3s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q2s32 = vmull_s16(d18s16, d1s16);
+    q3s32 = vmull_s16(d19s16, d1s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q13s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+    d14s16 = vqrshrn_n_s32(q2s32, 14);
+    d15s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q13s32, 14);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d0s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d0s16);
+    q3s32 = vmull_s16(d17s16, d0s16);
+    q13s32 = vmull_s16(d16s16, d0s16);
+    q15s32 = vmull_s16(d17s16, d0s16);
+
+    q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+    q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+    d0s16 = vdup_n_s16(cospi_24_64);
+    d1s16 = vdup_n_s16(cospi_8_64);
+
+    d18s16 = vqrshrn_n_s32(q2s32, 14);
+    d19s16 = vqrshrn_n_s32(q3s32, 14);
+    d22s16 = vqrshrn_n_s32(q13s32, 14);
+    d23s16 = vqrshrn_n_s32(q15s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q2s32 = vmull_s16(d20s16, d0s16);
+    q3s32 = vmull_s16(d21s16, d0s16);
+    q8s32 = vmull_s16(d20s16, d1s16);
+    q12s32 = vmull_s16(d21s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+    q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+    d26s16 = vqrshrn_n_s32(q2s32, 14);
+    d27s16 = vqrshrn_n_s32(q3s32, 14);
+    d30s16 = vqrshrn_n_s32(q8s32, 14);
+    d31s16 = vqrshrn_n_s32(q12s32, 14);
+    *q13s16 = vcombine_s16(d26s16, d27s16);
+    *q15s16 = vcombine_s16(d30s16, d31s16);
+
+    q0s16 = vaddq_s16(*q9s16, *q15s16);
+    q1s16 = vaddq_s16(*q11s16, *q13s16);
+    q2s16 = vsubq_s16(*q11s16, *q13s16);
+    q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+    *q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    *q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16 = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32 = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    *q8s16 = vaddq_s16(q0s16, q7s16);
+    *q9s16 = vaddq_s16(q1s16, q6s16);
+    *q10s16 = vaddq_s16(q2s16, q5s16);
+    *q11s16 = vaddq_s16(q3s16, q4s16);
+    *q12s16 = vsubq_s16(q3s16, q4s16);
+    *q13s16 = vsubq_s16(q2s16, q5s16);
+    *q14s16 = vsubq_s16(q1s16, q6s16);
+    *q15s16 = vsubq_s16(q0s16, q7s16);
+    return;
+}
+
+void vp9_idct8x8_64_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 16);
+    q11s16 = vld1q_s16(input + 24);
+    q12s16 = vld1q_s16(input + 32);
+    q13s16 = vld1q_s16(input + 40);
+    q14s16 = vld1q_s16(input + 48);
+    q15s16 = vld1q_s16(input + 56);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    d1 = d2 = dest;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+
+    q8s16 = q12s16;
+    q9s16 = q13s16;
+    q10s16 = q14s16;
+    q11s16 = q15s16;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    return;
+}
+
+void vp9_idct8x8_12_add_neon(
+        int16_t *input,
+        uint8_t *dest,
+        int dest_stride) {
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
+    int16x4_t d26s16, d27s16, d28s16, d29s16;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+    int32x4_t q9s32, q10s32, q11s32, q12s32;
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 16);
+    q11s16 = vld1q_s16(input + 24);
+    q12s16 = vld1q_s16(input + 32);
+    q13s16 = vld1q_s16(input + 40);
+    q14s16 = vld1q_s16(input + 48);
+    q15s16 = vld1q_s16(input + 56);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    // First transform rows
+    // stage 1
+    q0s16 = vdupq_n_s16(cospi_28_64 * 2);
+    q1s16 = vdupq_n_s16(cospi_4_64 * 2);
+
+    q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+
+    q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
+
+    q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+    q1s16 = vdupq_n_s16(cospi_12_64 * 2);
+
+    q5s16 = vqrdmulhq_s16(q11s16, q0s16);
+
+    q0s16 = vdupq_n_s16(cospi_16_64 * 2);
+
+    q6s16 = vqrdmulhq_s16(q11s16, q1s16);
+
+    // stage 2 & stage 3 - even half
+    q1s16 = vdupq_n_s16(cospi_24_64 * 2);
+
+    q9s16 = vqrdmulhq_s16(q8s16, q0s16);
+
+    q0s16 = vdupq_n_s16(cospi_8_64 * 2);
+
+    q13s16 = vqrdmulhq_s16(q10s16, q1s16);
+
+    q15s16 = vqrdmulhq_s16(q10s16, q0s16);
+
+    // stage 3 -odd half
+    q0s16 = vaddq_s16(q9s16, q15s16);
+    q1s16 = vaddq_s16(q9s16, q13s16);
+    q2s16 = vsubq_s16(q9s16, q13s16);
+    q3s16 = vsubq_s16(q9s16, q15s16);
+
+    // stage 2 - odd half
+    q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16 = vaddq_s16(q4s16, q5s16);
+    q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16 = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+    q9s32 = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32 = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    // stage 4
+    q8s16 = vaddq_s16(q0s16, q7s16);
+    q9s16 = vaddq_s16(q1s16, q6s16);
+    q10s16 = vaddq_s16(q2s16, q5s16);
+    q11s16 = vaddq_s16(q3s16, q4s16);
+    q12s16 = vsubq_s16(q3s16, q4s16);
+    q13s16 = vsubq_s16(q2s16, q5s16);
+    q14s16 = vsubq_s16(q1s16, q6s16);
+    q15s16 = vsubq_s16(q0s16, q7s16);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+               &q12s16, &q13s16, &q14s16, &q15s16);
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    d1 = d2 = dest;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+
+    q8s16 = q12s16;
+    q9s16 = q13s16;
+    q10s16 = q14s16;
+    q11s16 = q15s16;
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                     vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                     vreinterpret_u8_u64(d1u64));
+    q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                      vreinterpret_u8_u64(d2u64));
+    q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                      vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon_asm.asm
new file mode 100644
index 000000000..ab5bb6920
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_idct8x8_add_neon_asm.asm
@@ -0,0 +1,519 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_idct8x8_64_add_neon|
+    EXPORT  |vp9_idct8x8_12_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Parallel 1D IDCT on all the columns of a 8x8 16bit data matrix which are
+    ; loaded in q8-q15. The output will be stored back into q8-q15 registers.
+    ; This macro will touch q0-q7 registers and use them as buffer during
+    ; calculation.
+    MACRO
+    IDCT8x8_1D
+    ; stage 1
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r4                    ; duplicate cospi_4_64
+    vdup.16         d2, r5                    ; duplicate cospi_12_64
+    vdup.16         d3, r6                    ; duplicate cospi_20_64
+
+    ; input[1] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; input[5] * cospi_12_64
+    vmull.s16       q5, d26, d2
+    vmull.s16       q6, d27, d2
+
+    ; input[1]*cospi_28_64-input[7]*cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q5, d22, d3
+    vmlsl.s16       q6, d23, d3
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q5, #14              ; >> 14
+    vqrshrn.s32     d11, q6, #14              ; >> 14
+
+    ; input[1] * cospi_4_64
+    vmull.s16       q2, d18, d1
+    vmull.s16       q3, d19, d1
+
+    ; input[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q13, d27, d3
+
+    ; input[1]*cospi_4_64+input[7]*cospi_28_64
+    vmlal.s16       q2, d30, d0
+    vmlal.s16       q3, d31, d0
+
+    ; input[5] * cospi_20_64 + input[3] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q13, d23, d2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d14, q2, #14              ; >> 14
+    vqrshrn.s32     d15, q3, #14              ; >> 14
+
+    ; stage 2 & stage 3 - even half
+    vdup.16         d0, r7                    ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q13, #14              ; >> 14
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q2, d16, d0
+    vmull.s16       q3, d17, d0
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q13, d16, d0
+    vmull.s16       q15, d17, d0
+
+    ; (input[0] + input[2]) * cospi_16_64
+    vmlal.s16       q2,  d24, d0
+    vmlal.s16       q3, d25, d0
+
+    ; (input[0] - input[2]) * cospi_16_64
+    vmlsl.s16       q13, d24, d0
+    vmlsl.s16       q15, d25, d0
+
+    vdup.16         d0, r8                    ; duplicate cospi_24_64
+    vdup.16         d1, r9                    ; duplicate cospi_8_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d18, q2, #14              ; >> 14
+    vqrshrn.s32     d19, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d22, q13, #14              ; >> 14
+    vqrshrn.s32     d23, q15, #14              ; >> 14
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    ; input[1] * cospi_24_64
+    vmull.s16       q2, d20, d0
+    vmull.s16       q3, d21, d0
+
+    ; input[1] * cospi_8_64
+    vmull.s16       q8, d20, d1
+    vmull.s16       q12, d21, d1
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q2, d28, d1
+    vmlsl.s16       q3, d29, d1
+
+    ; input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q8, d28, d0
+    vmlal.s16       q12, d29, d0
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d26, q2, #14              ; >> 14
+    vqrshrn.s32     d27, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d30, q8, #14              ; >> 14
+    vqrshrn.s32     d31, q12, #14              ; >> 14
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q11, q13              ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q11, q13              ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+    MEND
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void vp9_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct8x8_64_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    IDCT8x8_1D
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |vp9_idct8x8_64_add_neon|
+
+;void vp9_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|vp9_idct8x8_12_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    ; stage 1
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(input[1] * cospi_28_64). vqrdmulh will do doubling
+    ; multiply and shift the result by 16 bits instead of 14 bits. So we need
+    ; to double the constants before multiplying to compensate this.
+    mov             r12, r3, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_28_64*2
+    mov             r12, r4, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_28_64)
+    vqrdmulh.s16    q4, q9, q0
+
+    mov             r12, r6, lsl #1
+    rsb             r12, #0
+    vdup.16         q0, r12                   ; duplicate -cospi_20_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_4_64)
+    vqrdmulh.s16    q7, q9, q1
+
+    mov             r12, r5, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_12_64*2
+
+    ; dct_const_round_shift(- input[3] * cospi_20_64)
+    vqrdmulh.s16    q5, q11, q0
+
+    mov             r12, r7, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_16_64*2
+
+    ; dct_const_round_shift(input[3] * cospi_12_64)
+    vqrdmulh.s16    q6, q11, q1
+
+    ; stage 2 & stage 3 - even half
+    mov             r12, r8, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_24_64*2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrdmulh.s16    q9, q8, q0
+
+    mov             r12, r9, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_8_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_24_64)
+    vqrdmulh.s16    q13, q10, q1
+
+    ; dct_const_round_shift(input[1] * cospi_8_64)
+    vqrdmulh.s16    q15, q10, q0
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q9, q13               ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q9, q13               ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |vp9_idct8x8_12_add_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
new file mode 100644
index 000000000..1761fada2
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
@@ -0,0 +1,248 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+static int16_t sinpi_1_9 = 0x14a3;
+static int16_t sinpi_2_9 = 0x26c9;
+static int16_t sinpi_3_9 = 0x3441;
+static int16_t sinpi_4_9 = 0x3b6c;
+static int16_t cospi_8_64 = 0x3b21;
+static int16_t cospi_16_64 = 0x2d41;
+static int16_t cospi_24_64 = 0x187e;
+
+static INLINE void TRANSPOSE4X4(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int32x4_t q8s32, q9s32;
+    int16x4x2_t d0x2s16, d1x2s16;
+    int32x4x2_t q0x2s32;
+
+    d0x2s16 = vtrn_s16(vget_low_s16(*q8s16), vget_high_s16(*q8s16));
+    d1x2s16 = vtrn_s16(vget_low_s16(*q9s16), vget_high_s16(*q9s16));
+
+    q8s32 = vreinterpretq_s32_s16(vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]));
+    q9s32 = vreinterpretq_s32_s16(vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]));
+    q0x2s32 = vtrnq_s32(q8s32, q9s32);
+
+    *q8s16 = vreinterpretq_s16_s32(q0x2s32.val[0]);
+    *q9s16 = vreinterpretq_s16_s32(q0x2s32.val[1]);
+    return;
+}
+
+static INLINE void GENERATE_COSINE_CONSTANTS(
+        int16x4_t *d0s16,
+        int16x4_t *d1s16,
+        int16x4_t *d2s16) {
+    *d0s16 = vdup_n_s16(cospi_8_64);
+    *d1s16 = vdup_n_s16(cospi_16_64);
+    *d2s16 = vdup_n_s16(cospi_24_64);
+    return;
+}
+
+static INLINE void GENERATE_SINE_CONSTANTS(
+        int16x4_t *d3s16,
+        int16x4_t *d4s16,
+        int16x4_t *d5s16,
+        int16x8_t *q3s16) {
+    *d3s16 = vdup_n_s16(sinpi_1_9);
+    *d4s16 = vdup_n_s16(sinpi_2_9);
+    *q3s16 = vdupq_n_s16(sinpi_3_9);
+    *d5s16 = vdup_n_s16(sinpi_4_9);
+    return;
+}
+
+static INLINE void IDCT4x4_1D(
+        int16x4_t *d0s16,
+        int16x4_t *d1s16,
+        int16x4_t *d2s16,
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
+    int16x4_t d26s16, d27s16, d28s16, d29s16;
+    int32x4_t q10s32, q13s32, q14s32, q15s32;
+    int16x8_t q13s16, q14s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+
+    d23s16 = vadd_s16(d16s16, d18s16);
+    d24s16 = vsub_s16(d16s16, d18s16);
+
+    q15s32 = vmull_s16(d17s16, *d2s16);
+    q10s32 = vmull_s16(d17s16, *d0s16);
+    q13s32 = vmull_s16(d23s16, *d1s16);
+    q14s32 = vmull_s16(d24s16, *d1s16);
+    q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
+    q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
+
+    d26s16 = vqrshrn_n_s32(q13s32, 14);
+    d27s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d28s16 = vqrshrn_n_s32(q10s32, 14);
+
+    q13s16 = vcombine_s16(d26s16, d27s16);
+    q14s16 = vcombine_s16(d28s16, d29s16);
+    *q8s16 = vaddq_s16(q13s16, q14s16);
+    *q9s16 = vsubq_s16(q13s16, q14s16);
+    *q9s16 = vcombine_s16(vget_high_s16(*q9s16),
+                          vget_low_s16(*q9s16));  // vswp
+    return;
+}
+
+static INLINE void IADST4x4_1D(
+        int16x4_t *d3s16,
+        int16x4_t *d4s16,
+        int16x4_t *d5s16,
+        int16x8_t *q3s16,
+        int16x8_t *q8s16,
+        int16x8_t *q9s16) {
+    int16x4_t d6s16, d16s16, d17s16, d18s16, d19s16;
+    int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+    d6s16 = vget_low_s16(*q3s16);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+
+    q10s32 = vmull_s16(*d3s16, d16s16);
+    q11s32 = vmull_s16(*d4s16, d16s16);
+    q12s32 = vmull_s16(d6s16, d17s16);
+    q13s32 = vmull_s16(*d5s16, d18s16);
+    q14s32 = vmull_s16(*d3s16, d18s16);
+    q15s32 = vmovl_s16(d16s16);
+    q15s32 = vaddw_s16(q15s32, d19s16);
+    q8s32  = vmull_s16(*d4s16, d19s16);
+    q15s32 = vsubw_s16(q15s32, d18s16);
+    q9s32  = vmull_s16(*d5s16, d19s16);
+
+    q10s32 = vaddq_s32(q10s32, q13s32);
+    q10s32 = vaddq_s32(q10s32, q8s32);
+    q11s32 = vsubq_s32(q11s32, q14s32);
+    q8s32  = vdupq_n_s32(sinpi_3_9);
+    q11s32 = vsubq_s32(q11s32, q9s32);
+    q15s32 = vmulq_s32(q15s32, q8s32);
+
+    q13s32 = vaddq_s32(q10s32, q12s32);
+    q10s32 = vaddq_s32(q10s32, q11s32);
+    q14s32 = vaddq_s32(q11s32, q12s32);
+    q10s32 = vsubq_s32(q10s32, q12s32);
+
+    d16s16 = vqrshrn_n_s32(q13s32, 14);
+    d17s16 = vqrshrn_n_s32(q14s32, 14);
+    d18s16 = vqrshrn_n_s32(q15s32, 14);
+    d19s16 = vqrshrn_n_s32(q10s32, 14);
+
+    *q8s16 = vcombine_s16(d16s16, d17s16);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+    return;
+}
+
+void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+    uint8x8_t d26u8, d27u8;
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
+    uint32x2_t d26u32, d27u32;
+    int16x8_t q3s16, q8s16, q9s16;
+    uint16x8_t q8u16, q9u16;
+
+    d26u32 = d27u32 = vdup_n_u32(0);
+
+    q8s16 = vld1q_s16(input);
+    q9s16 = vld1q_s16(input + 8);
+
+    TRANSPOSE4X4(&q8s16, &q9s16);
+
+    switch (tx_type) {
+      case 0:  // idct_idct is not supported. Fall back to C
+        vp9_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
+        return;
+        break;
+      case 1:  // iadst_idct
+        // generate constants
+        GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+        break;
+      case 2:  // idct_iadst
+        // generate constantsyy
+        GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+        break;
+      case 3:  // iadst_iadst
+        // generate constants
+        GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+        // first transform rows
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+        // transpose the matrix
+        TRANSPOSE4X4(&q8s16, &q9s16);
+
+        // then transform columns
+        IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+        break;
+      default:  // iadst_idct
+        assert(0);
+        break;
+    }
+
+    q8s16 = vrshrq_n_s16(q8s16, 4);
+    q9s16 = vrshrq_n_s16(q9s16, 4);
+
+    d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
+    dest += dest_stride;
+    d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
+    dest += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
+    dest += dest_stride;
+    d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
+
+    d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
+    dest -= dest_stride;
+    vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c
new file mode 100644
index 000000000..04b342c3d
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_iht8x8_add_neon.c
@@ -0,0 +1,624 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+static int16_t cospi_2_64 = 16305;
+static int16_t cospi_4_64 = 16069;
+static int16_t cospi_6_64 = 15679;
+static int16_t cospi_8_64 = 15137;
+static int16_t cospi_10_64 = 14449;
+static int16_t cospi_12_64 = 13623;
+static int16_t cospi_14_64 = 12665;
+static int16_t cospi_16_64 = 11585;
+static int16_t cospi_18_64 = 10394;
+static int16_t cospi_20_64 = 9102;
+static int16_t cospi_22_64 = 7723;
+static int16_t cospi_24_64 = 6270;
+static int16_t cospi_26_64 = 4756;
+static int16_t cospi_28_64 = 3196;
+static int16_t cospi_30_64 = 1606;
+
+static INLINE void TRANSPOSE8X8(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+    int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    *q8s16  = vcombine_s16(d16s16, d24s16);  // vswp d17, d24
+    *q9s16  = vcombine_s16(d18s16, d26s16);  // vswp d19, d26
+    *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    *q12s16 = vcombine_s16(d17s16, d25s16);
+    *q13s16 = vcombine_s16(d19s16, d27s16);
+    *q14s16 = vcombine_s16(d21s16, d29s16);
+    *q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q8s16),
+                        vreinterpretq_s32_s16(*q10s16));
+    q1x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q9s16),
+                        vreinterpretq_s32_s16(*q11s16));
+    q2x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q12s16),
+                        vreinterpretq_s32_s16(*q14s16));
+    q3x2s32 = vtrnq_s32(vreinterpretq_s32_s16(*q13s16),
+                        vreinterpretq_s32_s16(*q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    *q8s16  = q0x2s16.val[0];
+    *q9s16  = q0x2s16.val[1];
+    *q10s16 = q1x2s16.val[0];
+    *q11s16 = q1x2s16.val[1];
+    *q12s16 = q2x2s16.val[0];
+    *q13s16 = q2x2s16.val[1];
+    *q14s16 = q3x2s16.val[0];
+    *q15s16 = q3x2s16.val[1];
+    return;
+}
+
+static INLINE void IDCT8x8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+    int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+    int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+    d0s16 = vdup_n_s16(cospi_28_64);
+    d1s16 = vdup_n_s16(cospi_4_64);
+    d2s16 = vdup_n_s16(cospi_12_64);
+    d3s16 = vdup_n_s16(cospi_20_64);
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    q2s32 = vmull_s16(d18s16, d0s16);
+    q3s32 = vmull_s16(d19s16, d0s16);
+    q5s32 = vmull_s16(d26s16, d2s16);
+    q6s32 = vmull_s16(d27s16, d2s16);
+
+    q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+    q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+    q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+    d8s16  = vqrshrn_n_s32(q2s32, 14);
+    d9s16  = vqrshrn_n_s32(q3s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    q4s16 = vcombine_s16(d8s16, d9s16);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+
+    q2s32 = vmull_s16(d18s16, d1s16);
+    q3s32 = vmull_s16(d19s16, d1s16);
+    q9s32 = vmull_s16(d26s16, d3s16);
+    q13s32 = vmull_s16(d27s16, d3s16);
+
+    q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+    q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+    q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+    d14s16 = vqrshrn_n_s32(q2s32, 14);
+    d15s16 = vqrshrn_n_s32(q3s32, 14);
+    d12s16 = vqrshrn_n_s32(q9s32, 14);
+    d13s16 = vqrshrn_n_s32(q13s32, 14);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+    q7s16 = vcombine_s16(d14s16, d15s16);
+
+    d0s16 = vdup_n_s16(cospi_16_64);
+
+    q2s32 = vmull_s16(d16s16, d0s16);
+    q3s32 = vmull_s16(d17s16, d0s16);
+    q13s32 = vmull_s16(d16s16, d0s16);
+    q15s32 = vmull_s16(d17s16, d0s16);
+
+    q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+    q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+    d0s16 = vdup_n_s16(cospi_24_64);
+    d1s16 = vdup_n_s16(cospi_8_64);
+
+    d18s16 = vqrshrn_n_s32(q2s32, 14);
+    d19s16 = vqrshrn_n_s32(q3s32, 14);
+    d22s16 = vqrshrn_n_s32(q13s32, 14);
+    d23s16 = vqrshrn_n_s32(q15s32, 14);
+    *q9s16  = vcombine_s16(d18s16, d19s16);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q2s32 = vmull_s16(d20s16, d0s16);
+    q3s32 = vmull_s16(d21s16, d0s16);
+    q8s32 = vmull_s16(d20s16, d1s16);
+    q12s32 = vmull_s16(d21s16, d1s16);
+
+    q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+    q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+    q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+    q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+    d26s16 = vqrshrn_n_s32(q2s32, 14);
+    d27s16 = vqrshrn_n_s32(q3s32, 14);
+    d30s16 = vqrshrn_n_s32(q8s32, 14);
+    d31s16 = vqrshrn_n_s32(q12s32, 14);
+    *q13s16 = vcombine_s16(d26s16, d27s16);
+    *q15s16 = vcombine_s16(d30s16, d31s16);
+
+    q0s16 = vaddq_s16(*q9s16, *q15s16);
+    q1s16 = vaddq_s16(*q11s16, *q13s16);
+    q2s16 = vsubq_s16(*q11s16, *q13s16);
+    q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+    *q13s16 = vsubq_s16(q4s16, q5s16);
+    q4s16   = vaddq_s16(q4s16, q5s16);
+    *q14s16 = vsubq_s16(q7s16, q6s16);
+    q7s16   = vaddq_s16(q7s16, q6s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+
+    d16s16 = vdup_n_s16(cospi_16_64);
+
+    q9s32  = vmull_s16(d28s16, d16s16);
+    q10s32 = vmull_s16(d29s16, d16s16);
+    q11s32 = vmull_s16(d28s16, d16s16);
+    q12s32 = vmull_s16(d29s16, d16s16);
+
+    q9s32  = vmlsl_s16(q9s32,  d26s16, d16s16);
+    q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+    q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+    q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+    d10s16 = vqrshrn_n_s32(q9s32, 14);
+    d11s16 = vqrshrn_n_s32(q10s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q12s32, 14);
+    q5s16 = vcombine_s16(d10s16, d11s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    *q8s16  = vaddq_s16(q0s16, q7s16);
+    *q9s16  = vaddq_s16(q1s16, q6s16);
+    *q10s16 = vaddq_s16(q2s16, q5s16);
+    *q11s16 = vaddq_s16(q3s16, q4s16);
+    *q12s16 = vsubq_s16(q3s16, q4s16);
+    *q13s16 = vsubq_s16(q2s16, q5s16);
+    *q14s16 = vsubq_s16(q1s16, q6s16);
+    *q15s16 = vsubq_s16(q0s16, q7s16);
+    return;
+}
+
+static INLINE void IADST8X8_1D(
+        int16x8_t *q8s16,
+        int16x8_t *q9s16,
+        int16x8_t *q10s16,
+        int16x8_t *q11s16,
+        int16x8_t *q12s16,
+        int16x8_t *q13s16,
+        int16x8_t *q14s16,
+        int16x8_t *q15s16) {
+    int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+    int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+    int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+    int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+    int16x8_t q2s16, q4s16, q5s16, q6s16;
+    int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q7s32, q8s32;
+    int32x4_t q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+    d16s16 = vget_low_s16(*q8s16);
+    d17s16 = vget_high_s16(*q8s16);
+    d18s16 = vget_low_s16(*q9s16);
+    d19s16 = vget_high_s16(*q9s16);
+    d20s16 = vget_low_s16(*q10s16);
+    d21s16 = vget_high_s16(*q10s16);
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    d26s16 = vget_low_s16(*q13s16);
+    d27s16 = vget_high_s16(*q13s16);
+    d28s16 = vget_low_s16(*q14s16);
+    d29s16 = vget_high_s16(*q14s16);
+    d30s16 = vget_low_s16(*q15s16);
+    d31s16 = vget_high_s16(*q15s16);
+
+    d14s16 = vdup_n_s16(cospi_2_64);
+    d15s16 = vdup_n_s16(cospi_30_64);
+
+    q1s32 = vmull_s16(d30s16, d14s16);
+    q2s32 = vmull_s16(d31s16, d14s16);
+    q3s32 = vmull_s16(d30s16, d15s16);
+    q4s32 = vmull_s16(d31s16, d15s16);
+
+    d30s16 = vdup_n_s16(cospi_18_64);
+    d31s16 = vdup_n_s16(cospi_14_64);
+
+    q1s32 = vmlal_s16(q1s32, d16s16, d15s16);
+    q2s32 = vmlal_s16(q2s32, d17s16, d15s16);
+    q3s32 = vmlsl_s16(q3s32, d16s16, d14s16);
+    q4s32 = vmlsl_s16(q4s32, d17s16, d14s16);
+
+    q5s32 = vmull_s16(d22s16, d30s16);
+    q6s32 = vmull_s16(d23s16, d30s16);
+    q7s32 = vmull_s16(d22s16, d31s16);
+    q8s32 = vmull_s16(d23s16, d31s16);
+
+    q5s32 = vmlal_s16(q5s32, d24s16, d31s16);
+    q6s32 = vmlal_s16(q6s32, d25s16, d31s16);
+    q7s32 = vmlsl_s16(q7s32, d24s16, d30s16);
+    q8s32 = vmlsl_s16(q8s32, d25s16, d30s16);
+
+    q11s32 = vaddq_s32(q1s32, q5s32);
+    q12s32 = vaddq_s32(q2s32, q6s32);
+    q1s32 = vsubq_s32(q1s32, q5s32);
+    q2s32 = vsubq_s32(q2s32, q6s32);
+
+    d22s16 = vqrshrn_n_s32(q11s32, 14);
+    d23s16 = vqrshrn_n_s32(q12s32, 14);
+    *q11s16 = vcombine_s16(d22s16, d23s16);
+
+    q12s32 = vaddq_s32(q3s32, q7s32);
+    q15s32 = vaddq_s32(q4s32, q8s32);
+    q3s32 = vsubq_s32(q3s32, q7s32);
+    q4s32 = vsubq_s32(q4s32, q8s32);
+
+    d2s16  = vqrshrn_n_s32(q1s32, 14);
+    d3s16  = vqrshrn_n_s32(q2s32, 14);
+    d24s16 = vqrshrn_n_s32(q12s32, 14);
+    d25s16 = vqrshrn_n_s32(q15s32, 14);
+    d6s16  = vqrshrn_n_s32(q3s32, 14);
+    d7s16  = vqrshrn_n_s32(q4s32, 14);
+    *q12s16 = vcombine_s16(d24s16, d25s16);
+
+    d0s16 = vdup_n_s16(cospi_10_64);
+    d1s16 = vdup_n_s16(cospi_22_64);
+    q4s32 = vmull_s16(d26s16, d0s16);
+    q5s32 = vmull_s16(d27s16, d0s16);
+    q2s32 = vmull_s16(d26s16, d1s16);
+    q6s32 = vmull_s16(d27s16, d1s16);
+
+    d30s16 = vdup_n_s16(cospi_26_64);
+    d31s16 = vdup_n_s16(cospi_6_64);
+
+    q4s32 = vmlal_s16(q4s32, d20s16, d1s16);
+    q5s32 = vmlal_s16(q5s32, d21s16, d1s16);
+    q2s32 = vmlsl_s16(q2s32, d20s16, d0s16);
+    q6s32 = vmlsl_s16(q6s32, d21s16, d0s16);
+
+    q0s32 = vmull_s16(d18s16, d30s16);
+    q13s32 = vmull_s16(d19s16, d30s16);
+
+    q0s32 = vmlal_s16(q0s32, d28s16, d31s16);
+    q13s32 = vmlal_s16(q13s32, d29s16, d31s16);
+
+    q10s32 = vmull_s16(d18s16, d31s16);
+    q9s32 = vmull_s16(d19s16, d31s16);
+
+    q10s32 = vmlsl_s16(q10s32, d28s16, d30s16);
+    q9s32 = vmlsl_s16(q9s32, d29s16, d30s16);
+
+    q14s32 = vaddq_s32(q2s32, q10s32);
+    q15s32 = vaddq_s32(q6s32, q9s32);
+    q2s32 = vsubq_s32(q2s32, q10s32);
+    q6s32 = vsubq_s32(q6s32, q9s32);
+
+    d28s16 = vqrshrn_n_s32(q14s32, 14);
+    d29s16 = vqrshrn_n_s32(q15s32, 14);
+    d4s16 = vqrshrn_n_s32(q2s32, 14);
+    d5s16 = vqrshrn_n_s32(q6s32, 14);
+    *q14s16 = vcombine_s16(d28s16, d29s16);
+
+    q9s32 = vaddq_s32(q4s32, q0s32);
+    q10s32 = vaddq_s32(q5s32, q13s32);
+    q4s32 = vsubq_s32(q4s32, q0s32);
+    q5s32 = vsubq_s32(q5s32, q13s32);
+
+    d30s16 = vdup_n_s16(cospi_8_64);
+    d31s16 = vdup_n_s16(cospi_24_64);
+
+    d18s16 = vqrshrn_n_s32(q9s32, 14);
+    d19s16 = vqrshrn_n_s32(q10s32, 14);
+    d8s16 = vqrshrn_n_s32(q4s32, 14);
+    d9s16 = vqrshrn_n_s32(q5s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q5s32 = vmull_s16(d2s16, d30s16);
+    q6s32 = vmull_s16(d3s16, d30s16);
+    q7s32 = vmull_s16(d2s16, d31s16);
+    q0s32 = vmull_s16(d3s16, d31s16);
+
+    q5s32 = vmlal_s16(q5s32, d6s16, d31s16);
+    q6s32 = vmlal_s16(q6s32, d7s16, d31s16);
+    q7s32 = vmlsl_s16(q7s32, d6s16, d30s16);
+    q0s32 = vmlsl_s16(q0s32, d7s16, d30s16);
+
+    q1s32 = vmull_s16(d4s16, d30s16);
+    q3s32 = vmull_s16(d5s16, d30s16);
+    q10s32 = vmull_s16(d4s16, d31s16);
+    q2s32 = vmull_s16(d5s16, d31s16);
+
+    q1s32 = vmlsl_s16(q1s32, d8s16, d31s16);
+    q3s32 = vmlsl_s16(q3s32, d9s16, d31s16);
+    q10s32 = vmlal_s16(q10s32, d8s16, d30s16);
+    q2s32 = vmlal_s16(q2s32, d9s16, d30s16);
+
+    *q8s16 = vaddq_s16(*q11s16, *q9s16);
+    *q11s16 = vsubq_s16(*q11s16, *q9s16);
+    q4s16 = vaddq_s16(*q12s16, *q14s16);
+    *q12s16 = vsubq_s16(*q12s16, *q14s16);
+
+    q14s32 = vaddq_s32(q5s32, q1s32);
+    q15s32 = vaddq_s32(q6s32, q3s32);
+    q5s32 = vsubq_s32(q5s32, q1s32);
+    q6s32 = vsubq_s32(q6s32, q3s32);
+
+    d18s16 = vqrshrn_n_s32(q14s32, 14);
+    d19s16 = vqrshrn_n_s32(q15s32, 14);
+    d10s16 = vqrshrn_n_s32(q5s32, 14);
+    d11s16 = vqrshrn_n_s32(q6s32, 14);
+    *q9s16 = vcombine_s16(d18s16, d19s16);
+
+    q1s32 = vaddq_s32(q7s32, q10s32);
+    q3s32 = vaddq_s32(q0s32, q2s32);
+    q7s32 = vsubq_s32(q7s32, q10s32);
+    q0s32 = vsubq_s32(q0s32, q2s32);
+
+    d28s16 = vqrshrn_n_s32(q1s32, 14);
+    d29s16 = vqrshrn_n_s32(q3s32, 14);
+    d14s16 = vqrshrn_n_s32(q7s32, 14);
+    d15s16 = vqrshrn_n_s32(q0s32, 14);
+    *q14s16 = vcombine_s16(d28s16, d29s16);
+
+    d30s16 = vdup_n_s16(cospi_16_64);
+
+    d22s16 = vget_low_s16(*q11s16);
+    d23s16 = vget_high_s16(*q11s16);
+    q2s32 = vmull_s16(d22s16, d30s16);
+    q3s32 = vmull_s16(d23s16, d30s16);
+    q13s32 = vmull_s16(d22s16, d30s16);
+    q1s32 = vmull_s16(d23s16, d30s16);
+
+    d24s16 = vget_low_s16(*q12s16);
+    d25s16 = vget_high_s16(*q12s16);
+    q2s32 = vmlal_s16(q2s32, d24s16, d30s16);
+    q3s32 = vmlal_s16(q3s32, d25s16, d30s16);
+    q13s32 = vmlsl_s16(q13s32, d24s16, d30s16);
+    q1s32 = vmlsl_s16(q1s32, d25s16, d30s16);
+
+    d4s16 = vqrshrn_n_s32(q2s32, 14);
+    d5s16 = vqrshrn_n_s32(q3s32, 14);
+    d24s16 = vqrshrn_n_s32(q13s32, 14);
+    d25s16 = vqrshrn_n_s32(q1s32, 14);
+    q2s16 = vcombine_s16(d4s16, d5s16);
+    *q12s16 = vcombine_s16(d24s16, d25s16);
+
+    q13s32 = vmull_s16(d10s16, d30s16);
+    q1s32 = vmull_s16(d11s16, d30s16);
+    q11s32 = vmull_s16(d10s16, d30s16);
+    q0s32 = vmull_s16(d11s16, d30s16);
+
+    q13s32 = vmlal_s16(q13s32, d14s16, d30s16);
+    q1s32 = vmlal_s16(q1s32, d15s16, d30s16);
+    q11s32 = vmlsl_s16(q11s32, d14s16, d30s16);
+    q0s32 = vmlsl_s16(q0s32, d15s16, d30s16);
+
+    d20s16 = vqrshrn_n_s32(q13s32, 14);
+    d21s16 = vqrshrn_n_s32(q1s32, 14);
+    d12s16 = vqrshrn_n_s32(q11s32, 14);
+    d13s16 = vqrshrn_n_s32(q0s32, 14);
+    *q10s16 = vcombine_s16(d20s16, d21s16);
+    q6s16 = vcombine_s16(d12s16, d13s16);
+
+    q5s16 = vdupq_n_s16(0);
+
+    *q9s16  = vsubq_s16(q5s16, *q9s16);
+    *q11s16 = vsubq_s16(q5s16, q2s16);
+    *q13s16 = vsubq_s16(q5s16, q6s16);
+    *q15s16 = vsubq_s16(q5s16, q4s16);
+    return;
+}
+
+void vp9_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+    int i;
+    uint8_t *d1, *d2;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8;
+    uint64x1_t d0u64, d1u64, d2u64, d3u64;
+    int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+    uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+    q8s16  = vld1q_s16(input);
+    q9s16  = vld1q_s16(input + 8);
+    q10s16 = vld1q_s16(input + 8 * 2);
+    q11s16 = vld1q_s16(input + 8 * 3);
+    q12s16 = vld1q_s16(input + 8 * 4);
+    q13s16 = vld1q_s16(input + 8 * 5);
+    q14s16 = vld1q_s16(input + 8 * 6);
+    q15s16 = vld1q_s16(input + 8 * 7);
+
+    TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                 &q12s16, &q13s16, &q14s16, &q15s16);
+
+    switch (tx_type) {
+      case 0:  // idct_idct is not supported. Fall back to C
+        vp9_iht8x8_64_add_c(input, dest, dest_stride, tx_type);
+        return;
+        break;
+      case 1:  // iadst_idct
+        // generate IDCT constants
+        // GENERATE_IDCT_CONSTANTS
+
+        // first transform rows
+        IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                   &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // then transform columns
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      case 2:  // idct_iadst
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // first transform rows
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // generate IDCT constants
+        // GENERATE_IDCT_CONSTANTS
+
+        // then transform columns
+        IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                   &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      case 3:  // iadst_iadst
+        // generate IADST constants
+        // GENERATE_IADST_CONSTANTS
+
+        // first transform rows
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // transpose the matrix
+        TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16,
+                     &q12s16, &q13s16, &q14s16, &q15s16);
+
+        // then transform columns
+        IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16,
+                    &q12s16, &q13s16, &q14s16, &q15s16);
+        break;
+      default:  // iadst_idct
+        assert(0);
+        break;
+    }
+
+    q8s16 = vrshrq_n_s16(q8s16, 5);
+    q9s16 = vrshrq_n_s16(q9s16, 5);
+    q10s16 = vrshrq_n_s16(q10s16, 5);
+    q11s16 = vrshrq_n_s16(q11s16, 5);
+    q12s16 = vrshrq_n_s16(q12s16, 5);
+    q13s16 = vrshrq_n_s16(q13s16, 5);
+    q14s16 = vrshrq_n_s16(q14s16, 5);
+    q15s16 = vrshrq_n_s16(q15s16, 5);
+
+    for (d1 = d2 = dest, i = 0; i < 2; i++) {
+        if (i != 0) {
+            q8s16 = q12s16;
+            q9s16 = q13s16;
+            q10s16 = q14s16;
+            q11s16 = q15s16;
+        }
+
+        d0u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d1u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d2u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+        d3u64 = vld1_u64((uint64_t *)d1);
+        d1 += dest_stride;
+
+        q8u16  = vaddw_u8(vreinterpretq_u16_s16(q8s16),
+                          vreinterpret_u8_u64(d0u64));
+        q9u16  = vaddw_u8(vreinterpretq_u16_s16(q9s16),
+                          vreinterpret_u8_u64(d1u64));
+        q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16),
+                          vreinterpret_u8_u64(d2u64));
+        q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16),
+                          vreinterpret_u8_u64(d3u64));
+
+        d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+        d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+        d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+        d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+        d2 += dest_stride;
+        vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+        d2 += dest_stride;
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
new file mode 100644
index 000000000..c69ee1009
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
@@ -0,0 +1,179 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+static INLINE void vp9_loop_filter_neon_16(
+        uint8x16_t qblimit,  // blimit
+        uint8x16_t qlimit,   // limit
+        uint8x16_t qthresh,  // thresh
+        uint8x16_t q3,       // p3
+        uint8x16_t q4,       // p2
+        uint8x16_t q5,       // p1
+        uint8x16_t q6,       // p0
+        uint8x16_t q7,       // q0
+        uint8x16_t q8,       // q1
+        uint8x16_t q9,       // q2
+        uint8x16_t q10,      // q3
+        uint8x16_t *q5r,     // p1
+        uint8x16_t *q6r,     // p0
+        uint8x16_t *q7r,     // q0
+        uint8x16_t *q8r) {   // q1
+    uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+    int16x8_t q2s16, q11s16;
+    uint16x8_t q4u16;
+    int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
+    int8x8_t d2s8, d3s8;
+
+    q11u8 = vabdq_u8(q3, q4);
+    q12u8 = vabdq_u8(q4, q5);
+    q13u8 = vabdq_u8(q5, q6);
+    q14u8 = vabdq_u8(q8, q7);
+    q3 = vabdq_u8(q9, q8);
+    q4 = vabdq_u8(q10, q9);
+
+    q11u8 = vmaxq_u8(q11u8, q12u8);
+    q12u8 = vmaxq_u8(q13u8, q14u8);
+    q3 = vmaxq_u8(q3, q4);
+    q15u8 = vmaxq_u8(q11u8, q12u8);
+
+    q9 = vabdq_u8(q6, q7);
+
+    // vp8_hevmask
+    q13u8 = vcgtq_u8(q13u8, qthresh);
+    q14u8 = vcgtq_u8(q14u8, qthresh);
+    q15u8 = vmaxq_u8(q15u8, q3);
+
+    q2u8 = vabdq_u8(q5, q8);
+    q9 = vqaddq_u8(q9, q9);
+
+    q15u8 = vcgeq_u8(qlimit, q15u8);
+
+    // vp8_filter() function
+    // convert to signed
+    q10 = vdupq_n_u8(0x80);
+    q8 = veorq_u8(q8, q10);
+    q7 = veorq_u8(q7, q10);
+    q6 = veorq_u8(q6, q10);
+    q5 = veorq_u8(q5, q10);
+
+    q2u8 = vshrq_n_u8(q2u8, 1);
+    q9 = vqaddq_u8(q9, q2u8);
+
+    q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                     vget_low_s8(vreinterpretq_s8_u8(q6)));
+    q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                      vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+    q9 = vcgeq_u8(qblimit, q9);
+
+    q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+                    vreinterpretq_s8_u8(q8));
+
+    q14u8 = vorrq_u8(q13u8, q14u8);
+
+    q4u16 = vdupq_n_u16(3);
+    q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+    q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+    q15u8 = vandq_u8(q15u8, q9);
+
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+    q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+    q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+    q4 = vdupq_n_u8(3);
+    q9 = vdupq_n_u8(4);
+    // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+    d2s8 = vqmovn_s16(q2s16);
+    d3s8 = vqmovn_s16(q11s16);
+    q1s8 = vcombine_s8(d2s8, d3s8);
+    q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+    q1s8 = vreinterpretq_s8_u8(q1u8);
+
+    q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
+    q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+    q2s8 = vshrq_n_s8(q2s8, 3);
+    q1s8 = vshrq_n_s8(q1s8, 3);
+
+    q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+    q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+    q1s8 = vrshrq_n_s8(q1s8, 1);
+    q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+    q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+    q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+    *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
+    *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8),  q10);
+    *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
+    *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
+    return;
+}
+
+void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p /* pitch */,
+                                    const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+    uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
+    uint8x16_t qblimit, qlimit, qthresh;
+    uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+
+    dblimit0 = vld1_u8(blimit0);
+    dlimit0 = vld1_u8(limit0);
+    dthresh0 = vld1_u8(thresh0);
+    dblimit1 = vld1_u8(blimit1);
+    dlimit1 = vld1_u8(limit1);
+    dthresh1 = vld1_u8(thresh1);
+    qblimit = vcombine_u8(dblimit0, dblimit1);
+    qlimit = vcombine_u8(dlimit0, dlimit1);
+    qthresh = vcombine_u8(dthresh0, dthresh1);
+
+    s -= (p << 2);
+
+    q3u8 = vld1q_u8(s);
+    s += p;
+    q4u8 = vld1q_u8(s);
+    s += p;
+    q5u8 = vld1q_u8(s);
+    s += p;
+    q6u8 = vld1q_u8(s);
+    s += p;
+    q7u8 = vld1q_u8(s);
+    s += p;
+    q8u8 = vld1q_u8(s);
+    s += p;
+    q9u8 = vld1q_u8(s);
+    s += p;
+    q10u8 = vld1q_u8(s);
+
+    vp9_loop_filter_neon_16(qblimit, qlimit, qthresh,
+                            q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8,
+                            &q5u8, &q6u8, &q7u8, &q8u8);
+
+    s -= (p * 5);
+    vst1q_u8(s, q5u8);
+    s += p;
+    vst1q_u8(s, q6u8);
+    s += p;
+    vst1q_u8(s, q7u8);
+    s += p;
+    vst1q_u8(s, q8u8);
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon_asm.asm
new file mode 100644
index 000000000..5b8ec2028
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_16_neon_asm.asm
@@ -0,0 +1,199 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_lpf_horizontal_4_dual_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
+;                                    const uint8_t *blimit0,
+;                                    const uint8_t *limit0,
+;                                    const uint8_t *thresh0,
+;                                    const uint8_t *blimit1,
+;                                    const uint8_t *limit1,
+;                                    const uint8_t *thresh1)
+; r0    uint8_t *s,
+; r1    int p,
+; r2    const uint8_t *blimit0,
+; r3    const uint8_t *limit0,
+; sp    const uint8_t *thresh0,
+; sp+4  const uint8_t *blimit1,
+; sp+8  const uint8_t *limit1,
+; sp+12 const uint8_t *thresh1,
+
+|vp9_lpf_horizontal_4_dual_neon| PROC
+    push        {lr}
+
+    ldr         r12, [sp, #4]              ; load thresh0
+    vld1.8      {d0}, [r2]                 ; load blimit0 to first half q
+    vld1.8      {d2}, [r3]                 ; load limit0 to first half q
+
+    add         r1, r1, r1                 ; double pitch
+    ldr         r2, [sp, #8]               ; load blimit1
+
+    vld1.8      {d4}, [r12]                ; load thresh0 to first half q
+
+    ldr         r3, [sp, #12]              ; load limit1
+    ldr         r12, [sp, #16]             ; load thresh1
+    vld1.8      {d1}, [r2]                 ; load blimit1 to 2nd half q
+
+    sub         r2, r0, r1, lsl #1         ; s[-4 * p]
+
+    vld1.8      {d3}, [r3]                 ; load limit1 to 2nd half q
+    vld1.8      {d5}, [r12]                ; load thresh1 to 2nd half q
+
+    vpush       {d8-d15}                   ; save neon registers
+
+    add         r3, r2, r1, lsr #1         ; s[-3 * p]
+
+    vld1.u8     {q3}, [r2@64], r1          ; p3
+    vld1.u8     {q4}, [r3@64], r1          ; p2
+    vld1.u8     {q5}, [r2@64], r1          ; p1
+    vld1.u8     {q6}, [r3@64], r1          ; p0
+    vld1.u8     {q7}, [r2@64], r1          ; q0
+    vld1.u8     {q8}, [r3@64], r1          ; q1
+    vld1.u8     {q9}, [r2@64]              ; q2
+    vld1.u8     {q10}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          vp9_loop_filter_neon_16
+
+    vst1.u8     {q5}, [r2@64], r1          ; store op1
+    vst1.u8     {q6}, [r3@64], r1          ; store op0
+    vst1.u8     {q7}, [r2@64], r1          ; store oq0
+    vst1.u8     {q8}, [r3@64], r1          ; store oq1
+
+    vpop        {d8-d15}                   ; restore neon registers
+
+    pop         {pc}
+    ENDP        ; |vp9_lpf_horizontal_4_dual_neon|
+
+; void vp9_loop_filter_neon_16();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. This function uses
+; registers d8-d15, so the calling function must save those registers.
+;
+; r0-r3, r12 PRESERVE
+; q0    blimit
+; q1    limit
+; q2    thresh
+; q3    p3
+; q4    p2
+; q5    p1
+; q6    p0
+; q7    q0
+; q8    q1
+; q9    q2
+; q10   q3
+;
+; Outputs:
+; q5    op1
+; q6    op0
+; q7    oq0
+; q8    oq1
+|vp9_loop_filter_neon_16| PROC
+
+    ; filter_mask
+    vabd.u8     q11, q3, q4                 ; m1 = abs(p3 - p2)
+    vabd.u8     q12, q4, q5                 ; m2 = abs(p2 - p1)
+    vabd.u8     q13, q5, q6                 ; m3 = abs(p1 - p0)
+    vabd.u8     q14, q8, q7                 ; m4 = abs(q1 - q0)
+    vabd.u8     q3, q9, q8                  ; m5 = abs(q2 - q1)
+    vabd.u8     q4, q10, q9                 ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     q11, q11, q12               ; m7 = max(m1, m2)
+    vmax.u8     q12, q13, q14               ; m8 = max(m3, m4)
+
+    vabd.u8     q9, q6, q7                  ; abs(p0 - q0)
+
+    vmax.u8     q3, q3, q4                  ; m9 = max(m5, m6)
+
+    vmov.u8     q10, #0x80
+
+    vmax.u8     q15, q11, q12               ; m10 = max(m7, m8)
+
+    vcgt.u8     q13, q13, q2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     q14, q14, q2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     q15, q15, q3                ; m11 = max(m10, m9)
+
+    vabd.u8     q2, q5, q8                  ; a = abs(p1 - q1)
+    vqadd.u8    q9, q9, q9                  ; b = abs(p0 - q0) * 2
+
+    veor        q7, q7, q10                 ; qs0
+
+    vcge.u8     q15, q1, q15                ; abs(m11) > limit
+
+    vshr.u8     q2, q2, #1                  ; a = a / 2
+    veor        q6, q6, q10                 ; ps0
+
+    veor        q5, q5, q10                 ; ps1
+    vqadd.u8    q9, q9, q2                  ; a = b + a
+
+    veor        q8, q8, q10                 ; qs1
+
+    vmov.u16    q4, #3
+
+    vsubl.s8    q2, d14, d12                ; ( qs0 - ps0)
+    vsubl.s8    q11, d15, d13
+
+    vcge.u8     q9, q0, q9                  ; a > blimit
+
+    vqsub.s8    q1, q5, q8                  ; filter = clamp(ps1-qs1)
+    vorr        q14, q13, q14               ; hev
+
+    vmul.i16    q2, q2, q4                  ; 3 * ( qs0 - ps0)
+    vmul.i16    q11, q11, q4
+
+    vand        q1, q1, q14                 ; filter &= hev
+    vand        q15, q15, q9                ; mask
+
+    vmov.u8     q4, #3
+
+    vaddw.s8    q2, q2, d2                  ; filter + 3 * (qs0 - ps0)
+    vaddw.s8    q11, q11, d3
+
+    vmov.u8     q9, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d2, q2
+    vqmovn.s16  d3, q11
+    vand        q1, q1, q15                 ; filter &= mask
+
+    vqadd.s8    q2, q1, q4                  ; filter2 = clamp(filter+3)
+    vqadd.s8    q1, q1, q9                  ; filter1 = clamp(filter+4)
+    vshr.s8     q2, q2, #3                  ; filter2 >>= 3
+    vshr.s8     q1, q1, #3                  ; filter1 >>= 3
+
+
+    vqadd.s8    q11, q6, q2                 ; u = clamp(ps0 + filter2)
+    vqsub.s8    q0, q7, q1                  ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    q1, q1, #1                  ; filter = ++filter1 >> 1
+
+    veor        q7, q0,  q10                ; *oq0 = u^0x80
+
+    vbic        q1, q1, q14                 ; filter &= ~hev
+
+    vqadd.s8    q13, q5, q1                 ; u = clamp(ps1 + filter)
+    vqsub.s8    q12, q8, q1                 ; u = clamp(qs1 - filter)
+
+    veor        q6, q11, q10                ; *op0 = u^0x80
+    veor        q5, q13, q10                ; *op1 = u^0x80
+    veor        q8, q12, q10                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |vp9_loop_filter_neon_16|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon.c
new file mode 100644
index 000000000..fd9db6187
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon.c
@@ -0,0 +1,274 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+static INLINE void vp9_loop_filter_neon(
+        uint8x8_t dblimit,    // flimit
+        uint8x8_t dlimit,     // limit
+        uint8x8_t dthresh,    // thresh
+        uint8x8_t d3u8,       // p3
+        uint8x8_t d4u8,       // p2
+        uint8x8_t d5u8,       // p1
+        uint8x8_t d6u8,       // p0
+        uint8x8_t d7u8,       // q0
+        uint8x8_t d16u8,      // q1
+        uint8x8_t d17u8,      // q2
+        uint8x8_t d18u8,      // q3
+        uint8x8_t *d4ru8,     // p1
+        uint8x8_t *d5ru8,     // p0
+        uint8x8_t *d6ru8,     // q0
+        uint8x8_t *d7ru8) {   // q1
+    uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
+    int16x8_t q12s16;
+    int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
+
+    d19u8 = vabd_u8(d3u8, d4u8);
+    d20u8 = vabd_u8(d4u8, d5u8);
+    d21u8 = vabd_u8(d5u8, d6u8);
+    d22u8 = vabd_u8(d16u8, d7u8);
+    d3u8  = vabd_u8(d17u8, d16u8);
+    d4u8  = vabd_u8(d18u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+    d20u8 = vmax_u8(d21u8, d22u8);
+    d3u8  = vmax_u8(d3u8,  d4u8);
+    d23u8 = vmax_u8(d19u8, d20u8);
+
+    d17u8 = vabd_u8(d6u8, d7u8);
+
+    d21u8 = vcgt_u8(d21u8, dthresh);
+    d22u8 = vcgt_u8(d22u8, dthresh);
+    d23u8 = vmax_u8(d23u8, d3u8);
+
+    d28u8 = vabd_u8(d5u8, d16u8);
+    d17u8 = vqadd_u8(d17u8, d17u8);
+
+    d23u8 = vcge_u8(dlimit, d23u8);
+
+    d18u8 = vdup_n_u8(0x80);
+    d5u8  = veor_u8(d5u8,  d18u8);
+    d6u8  = veor_u8(d6u8,  d18u8);
+    d7u8  = veor_u8(d7u8,  d18u8);
+    d16u8 = veor_u8(d16u8, d18u8);
+
+    d28u8 = vshr_n_u8(d28u8, 1);
+    d17u8 = vqadd_u8(d17u8, d28u8);
+
+    d19u8 = vdup_n_u8(3);
+
+    d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8),
+                    vreinterpret_s8_u8(d6u8));
+
+    d17u8 = vcge_u8(dblimit, d17u8);
+
+    d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8),
+                     vreinterpret_s8_u8(d16u8));
+
+    d22u8 = vorr_u8(d21u8, d22u8);
+
+    q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+    d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
+    d23u8 = vand_u8(d23u8, d17u8);
+
+    q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
+
+    d17u8 = vdup_n_u8(4);
+
+    d27s8 = vqmovn_s16(q12s16);
+    d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
+    d27s8 = vreinterpret_s8_u8(d27u8);
+
+    d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
+    d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
+    d28s8 = vshr_n_s8(d28s8, 3);
+    d27s8 = vshr_n_s8(d27s8, 3);
+
+    d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
+    d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
+
+    d27s8 = vrshr_n_s8(d27s8, 1);
+    d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
+
+    d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
+    d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
+
+    *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
+    *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
+    *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
+    *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
+    return;
+}
+
+void vp9_lpf_horizontal_4_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char *blimit,
+        unsigned char *limit,
+        unsigned char *thresh,
+        int count) {
+    int i;
+    uint8_t *s, *psrc;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+
+    if (count == 0)  // end_vp9_lf_h_edge
+        return;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    psrc = src - (pitch << 2);
+    for (i = 0; i < count; i++) {
+        s = psrc + i * 8;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        vp9_loop_filter_neon(dblimit, dlimit, dthresh,
+                             d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                             &d4u8, &d5u8, &d6u8, &d7u8);
+
+        s -= (pitch * 5);
+        vst1_u8(s, d4u8);
+        s += pitch;
+        vst1_u8(s, d5u8);
+        s += pitch;
+        vst1_u8(s, d6u8);
+        s += pitch;
+        vst1_u8(s, d7u8);
+    }
+    return;
+}
+
+void vp9_lpf_vertical_4_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char *blimit,
+        unsigned char *limit,
+        unsigned char *thresh,
+        int count) {
+    int i, pitch8;
+    uint8_t *s;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+    uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+    uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+    uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+    uint8x8x4_t d4Result;
+
+    if (count == 0)  // end_vp9_lf_h_edge
+        return;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    pitch8 = pitch * 8;
+    for (i = 0; i < count; i++, src += pitch8) {
+        s = src - (i + 1) * 4;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+                      vreinterpret_u32_u8(d7u8));
+        d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+                      vreinterpret_u32_u8(d16u8));
+        d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+                      vreinterpret_u32_u8(d17u8));
+        d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+                      vreinterpret_u32_u8(d18u8));
+
+        d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                          vreinterpret_u16_u32(d2tmp2.val[0]));
+        d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                          vreinterpret_u16_u32(d2tmp3.val[0]));
+        d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                          vreinterpret_u16_u32(d2tmp2.val[1]));
+        d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                          vreinterpret_u16_u32(d2tmp3.val[1]));
+
+        d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                         vreinterpret_u8_u16(d2tmp5.val[0]));
+        d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                         vreinterpret_u8_u16(d2tmp5.val[1]));
+        d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                          vreinterpret_u8_u16(d2tmp7.val[0]));
+        d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                          vreinterpret_u8_u16(d2tmp7.val[1]));
+
+        d3u8 = d2tmp8.val[0];
+        d4u8 = d2tmp8.val[1];
+        d5u8 = d2tmp9.val[0];
+        d6u8 = d2tmp9.val[1];
+        d7u8 = d2tmp10.val[0];
+        d16u8 = d2tmp10.val[1];
+        d17u8 = d2tmp11.val[0];
+        d18u8 = d2tmp11.val[1];
+
+        vp9_loop_filter_neon(dblimit, dlimit, dthresh,
+                             d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                             &d4u8, &d5u8, &d6u8, &d7u8);
+
+        d4Result.val[0] = d4u8;
+        d4Result.val[1] = d5u8;
+        d4Result.val[2] = d6u8;
+        d4Result.val[3] = d7u8;
+
+        src -= 2;
+        vst4_lane_u8(src, d4Result, 0);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 1);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 2);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 3);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 4);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 5);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 6);
+        src += pitch;
+        vst4_lane_u8(src, d4Result, 7);
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon_asm.asm
new file mode 100644
index 000000000..7738e0d3a
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_4_neon_asm.asm
@@ -0,0 +1,277 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_lpf_horizontal_4_neon|
+    EXPORT  |vp9_lpf_vertical_4_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_horizontal_4_neon(uint8_t *s,
+;                                int p /* pitch */,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh,
+;                                int count)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; sp+4  int count
+|vp9_lpf_horizontal_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r12, [sp, #8]              ; load count
+    ldr         r2, [sp, #4]               ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    cmp         r12, #0
+    beq         end_vp9_lf_h_edge
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+count_lf_h_loop
+    sub         r2, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r3, r2, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r2@64], r1          ; p3
+    vld1.u8     {d4}, [r3@64], r1          ; p2
+    vld1.u8     {d5}, [r2@64], r1          ; p1
+    vld1.u8     {d6}, [r3@64], r1          ; p0
+    vld1.u8     {d7}, [r2@64], r1          ; q0
+    vld1.u8     {d16}, [r3@64], r1         ; q1
+    vld1.u8     {d17}, [r2@64]             ; q2
+    vld1.u8     {d18}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          vp9_loop_filter_neon
+
+    vst1.u8     {d4}, [r2@64], r1          ; store op1
+    vst1.u8     {d5}, [r3@64], r1          ; store op0
+    vst1.u8     {d6}, [r2@64], r1          ; store oq0
+    vst1.u8     {d7}, [r3@64], r1          ; store oq1
+
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         count_lf_h_loop
+
+end_vp9_lf_h_edge
+    pop         {pc}
+    ENDP        ; |vp9_lpf_horizontal_4_neon|
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_vertical_4_neon(uint8_t *s,
+;                              int p /* pitch */,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh,
+;                              int count)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; sp+4  int count
+|vp9_lpf_vertical_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    ldr         r12, [sp, #8]             ; load count
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #4]              ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+    cmp         r12, #0
+    beq         end_vp9_lf_v_edge
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+count_lf_v_loop
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    bl          vp9_loop_filter_neon
+
+    sub         r0, r0, #2
+
+    ;store op1, op0, oq0, oq1
+    vst4.8      {d4[0], d5[0], d6[0], d7[0]}, [r0], r1
+    vst4.8      {d4[1], d5[1], d6[1], d7[1]}, [r0], r1
+    vst4.8      {d4[2], d5[2], d6[2], d7[2]}, [r0], r1
+    vst4.8      {d4[3], d5[3], d6[3], d7[3]}, [r0], r1
+    vst4.8      {d4[4], d5[4], d6[4], d7[4]}, [r0], r1
+    vst4.8      {d4[5], d5[5], d6[5], d7[5]}, [r0], r1
+    vst4.8      {d4[6], d5[6], d6[6], d7[6]}, [r0], r1
+    vst4.8      {d4[7], d5[7], d6[7], d7[7]}, [r0]
+
+    add         r0, r0, r1, lsl #3         ; s += pitch * 8
+    subs        r12, r12, #1
+    subne       r2, r0, #4                 ; move s pointer down by 4 columns
+    bne         count_lf_v_loop
+
+end_vp9_lf_v_edge
+    pop         {pc}
+    ENDP        ; |vp9_lpf_vertical_4_neon|
+
+; void vp9_loop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d4    op1
+; d5    op0
+; d6    oq0
+; d7    oq1
+|vp9_loop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                 ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                 ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                 ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7                ; m4 = abs(q1 - q0)
+    vabd.u8     d3, d17, d16                ; m5 = abs(q2 - q1)
+    vabd.u8     d4, d18, d17                ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20               ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22               ; m2 = max(m3, m4)
+
+    vabd.u8     d17, d6, d7                 ; abs(p0 - q0)
+
+    vmax.u8     d3, d3, d4                  ; m3 = max(m5, m6)
+
+    vmov.u8     d18, #0x80
+
+    vmax.u8     d23, d19, d20               ; m1 = max(m1, m2)
+
+    ; hevmask
+    vcgt.u8     d21, d21, d2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     d23, d23, d3                ; m1 = max(m1, m3)
+
+    vabd.u8     d28, d5, d16                ; a = abs(p1 - q1)
+    vqadd.u8    d17, d17, d17               ; b = abs(p0 - q0) * 2
+
+    veor        d7, d7, d18                 ; qs0
+
+    vcge.u8     d23, d1, d23                ; abs(m1) > limit
+
+    ; filter() function
+    ; convert to signed
+
+    vshr.u8     d28, d28, #1                ; a = a / 2
+    veor        d6, d6, d18                 ; ps0
+
+    veor        d5, d5, d18                 ; ps1
+    vqadd.u8    d17, d17, d28               ; a = b + a
+
+    veor        d16, d16, d18               ; qs1
+
+    vmov.u8     d19, #3
+
+    vsub.s8     d28, d7, d6                 ; ( qs0 - ps0)
+
+    vcge.u8     d17, d0, d17                ; a > blimit
+
+    vqsub.s8    d27, d5, d16                ; filter = clamp(ps1-qs1)
+    vorr        d22, d21, d22               ; hevmask
+
+    vmull.s8    q12, d28, d19               ; 3 * ( qs0 - ps0)
+
+    vand        d27, d27, d22               ; filter &= hev
+    vand        d23, d23, d17               ; filter_mask
+
+    vaddw.s8    q12, q12, d27               ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d17, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d27, q12
+
+    vand        d27, d27, d23               ; filter &= mask
+
+    vqadd.s8    d28, d27, d19               ; filter2 = clamp(filter+3)
+    vqadd.s8    d27, d27, d17               ; filter1 = clamp(filter+4)
+    vshr.s8     d28, d28, #3                ; filter2 >>= 3
+    vshr.s8     d27, d27, #3                ; filter1 >>= 3
+
+    vqadd.s8    d19, d6, d28                ; u = clamp(ps0 + filter2)
+    vqsub.s8    d26, d7, d27                ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    d27, d27, #1                ; filter = ++filter1 >> 1
+
+    veor        d6, d26, d18                ; *oq0 = u^0x80
+
+    vbic        d27, d27, d22               ; filter &= ~hev
+
+    vqadd.s8    d21, d5, d27                ; u = clamp(ps1 + filter)
+    vqsub.s8    d20, d16, d27               ; u = clamp(qs1 - filter)
+
+    veor        d5, d19, d18                ; *op0 = u^0x80
+    veor        d4, d21, d18                ; *op1 = u^0x80
+    veor        d7, d20, d18                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |vp9_loop_filter_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon.c
new file mode 100644
index 000000000..33068a8a2
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon.c
@@ -0,0 +1,453 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_config.h"
+
+static INLINE void vp9_mbloop_filter_neon(
+        uint8x8_t dblimit,   // mblimit
+        uint8x8_t dlimit,    // limit
+        uint8x8_t dthresh,   // thresh
+        uint8x8_t d3u8,      // p2
+        uint8x8_t d4u8,      // p2
+        uint8x8_t d5u8,      // p1
+        uint8x8_t d6u8,      // p0
+        uint8x8_t d7u8,      // q0
+        uint8x8_t d16u8,     // q1
+        uint8x8_t d17u8,     // q2
+        uint8x8_t d18u8,     // q3
+        uint8x8_t *d0ru8,    // p1
+        uint8x8_t *d1ru8,    // p1
+        uint8x8_t *d2ru8,    // p0
+        uint8x8_t *d3ru8,    // q0
+        uint8x8_t *d4ru8,    // q1
+        uint8x8_t *d5ru8) {  // q1
+    uint32_t flat;
+    uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
+    uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+    int16x8_t q15s16;
+    uint16x8_t q10u16, q14u16;
+    int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
+
+    d19u8 = vabd_u8(d3u8, d4u8);
+    d20u8 = vabd_u8(d4u8, d5u8);
+    d21u8 = vabd_u8(d5u8, d6u8);
+    d22u8 = vabd_u8(d16u8, d7u8);
+    d23u8 = vabd_u8(d17u8, d16u8);
+    d24u8 = vabd_u8(d18u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+    d20u8 = vmax_u8(d21u8, d22u8);
+
+    d25u8 = vabd_u8(d6u8, d4u8);
+
+    d23u8 = vmax_u8(d23u8, d24u8);
+
+    d26u8 = vabd_u8(d7u8, d17u8);
+
+    d19u8 = vmax_u8(d19u8, d20u8);
+
+    d24u8 = vabd_u8(d6u8, d7u8);
+    d27u8 = vabd_u8(d3u8, d6u8);
+    d28u8 = vabd_u8(d18u8, d7u8);
+
+    d19u8 = vmax_u8(d19u8, d23u8);
+
+    d23u8 = vabd_u8(d5u8, d16u8);
+    d24u8 = vqadd_u8(d24u8, d24u8);
+
+
+    d19u8 = vcge_u8(dlimit, d19u8);
+
+
+    d25u8 = vmax_u8(d25u8, d26u8);
+    d26u8 = vmax_u8(d27u8, d28u8);
+
+    d23u8 = vshr_n_u8(d23u8, 1);
+
+    d25u8 = vmax_u8(d25u8, d26u8);
+
+    d24u8 = vqadd_u8(d24u8, d23u8);
+
+    d20u8 = vmax_u8(d20u8, d25u8);
+
+    d23u8 = vdup_n_u8(1);
+    d24u8 = vcge_u8(dblimit, d24u8);
+
+    d21u8 = vcgt_u8(d21u8, dthresh);
+
+    d20u8 = vcge_u8(d23u8, d20u8);
+
+    d19u8 = vand_u8(d19u8, d24u8);
+
+    d23u8 = vcgt_u8(d22u8, dthresh);
+
+    d20u8 = vand_u8(d20u8, d19u8);
+
+    d22u8 = vdup_n_u8(0x80);
+
+    d23u8 = vorr_u8(d21u8, d23u8);
+
+    q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8),
+                          vreinterpret_u16_u8(d21u8));
+
+    d30u8 = vshrn_n_u16(q10u16, 4);
+    flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
+
+    if (flat == 0xffffffff) {  // Check for all 1's, power_branch_only
+        d27u8 = vdup_n_u8(3);
+        d21u8 = vdup_n_u8(2);
+        q14u16 = vaddl_u8(d6u8, d7u8);
+        q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+        q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        *d0ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+        *d1ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        *d2ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d3ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vsubw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d4ru8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vsubw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+        *d5ru8 = vqrshrn_n_u16(q14u16, 3);
+    } else {
+        d21u8 = veor_u8(d7u8,  d22u8);
+        d24u8 = veor_u8(d6u8,  d22u8);
+        d25u8 = veor_u8(d5u8,  d22u8);
+        d26u8 = veor_u8(d16u8, d22u8);
+
+        d27u8 = vdup_n_u8(3);
+
+        d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
+        d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
+
+        q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
+
+        d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+        q15s16 = vaddw_s8(q15s16, d29s8);
+
+        d29u8 = vdup_n_u8(4);
+
+        d28s8 = vqmovn_s16(q15s16);
+
+        d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+        d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
+        d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
+        d30s8 = vshr_n_s8(d30s8, 3);
+        d29s8 = vshr_n_s8(d29s8, 3);
+
+        d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
+        d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
+
+        d29s8 = vrshr_n_s8(d29s8, 1);
+        d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+        d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
+        d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
+
+        if (flat == 0) {  // filter_branch_only
+            *d0ru8 = d4u8;
+            *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+            *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+            *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+            *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+            *d5ru8 = d17u8;
+            return;
+        }
+
+        d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+        d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+        d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+        d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+
+        d23u8 = vdup_n_u8(2);
+        q14u16 = vaddl_u8(d6u8, d7u8);
+        q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+        q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
+
+        d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
+
+        q14u16 = vaddw_u8(q14u16, d5u8);
+
+        d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
+
+        d30u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vaddw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+
+        d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
+
+        d31u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vaddw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+
+        *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
+
+        d23u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d3u8);
+        q14u16 = vsubw_u8(q14u16, d6u8);
+        q14u16 = vaddw_u8(q14u16, d7u8);
+
+        *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
+
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
+
+        d22u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d4u8);
+        q14u16 = vsubw_u8(q14u16, d7u8);
+        q14u16 = vaddw_u8(q14u16, d16u8);
+
+        d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
+
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
+
+        d6u8 = vqrshrn_n_u16(q14u16, 3);
+
+        q14u16 = vsubw_u8(q14u16, d5u8);
+        q14u16 = vsubw_u8(q14u16, d16u8);
+        q14u16 = vaddw_u8(q14u16, d17u8);
+        q14u16 = vaddw_u8(q14u16, d18u8);
+
+        d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
+
+        d7u8 = vqrshrn_n_u16(q14u16, 3);
+
+        *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
+        *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
+        *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
+    }
+    return;
+}
+
+void vp9_lpf_horizontal_8_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char *blimit,
+        unsigned char *limit,
+        unsigned char *thresh,
+        int count) {
+    int i;
+    uint8_t *s, *psrc;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    uint8x8_t d16u8, d17u8, d18u8;
+
+    if (count == 0)  // end_vp9_mblf_h_edge
+        return;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    psrc = src - (pitch << 2);
+    for (i = 0; i < count; i++) {
+        s = psrc + i * 8;
+
+        d3u8  = vld1_u8(s);
+        s += pitch;
+        d4u8  = vld1_u8(s);
+        s += pitch;
+        d5u8  = vld1_u8(s);
+        s += pitch;
+        d6u8  = vld1_u8(s);
+        s += pitch;
+        d7u8  = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
+                             d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                             &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+        s -= (pitch * 6);
+        vst1_u8(s, d0u8);
+        s += pitch;
+        vst1_u8(s, d1u8);
+        s += pitch;
+        vst1_u8(s, d2u8);
+        s += pitch;
+        vst1_u8(s, d3u8);
+        s += pitch;
+        vst1_u8(s, d4u8);
+        s += pitch;
+        vst1_u8(s, d5u8);
+    }
+    return;
+}
+
+void vp9_lpf_vertical_8_neon(
+        unsigned char *src,
+        int pitch,
+        unsigned char *blimit,
+        unsigned char *limit,
+        unsigned char *thresh,
+        int count) {
+    int i;
+    uint8_t *s;
+    uint8x8_t dblimit, dlimit, dthresh;
+    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+    uint8x8_t d16u8, d17u8, d18u8;
+    uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+    uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+    uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+    uint8x8x4_t d4Result;
+    uint8x8x2_t d2Result;
+
+    if (count == 0)
+        return;
+
+    dblimit = vld1_u8(blimit);
+    dlimit = vld1_u8(limit);
+    dthresh = vld1_u8(thresh);
+
+    for (i = 0; i < count; i++) {
+        s = src + (i * (pitch << 3)) - 4;
+
+        d3u8 = vld1_u8(s);
+        s += pitch;
+        d4u8 = vld1_u8(s);
+        s += pitch;
+        d5u8 = vld1_u8(s);
+        s += pitch;
+        d6u8 = vld1_u8(s);
+        s += pitch;
+        d7u8 = vld1_u8(s);
+        s += pitch;
+        d16u8 = vld1_u8(s);
+        s += pitch;
+        d17u8 = vld1_u8(s);
+        s += pitch;
+        d18u8 = vld1_u8(s);
+
+        d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+                          vreinterpret_u32_u8(d7u8));
+        d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+                          vreinterpret_u32_u8(d16u8));
+        d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+                          vreinterpret_u32_u8(d17u8));
+        d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+                          vreinterpret_u32_u8(d18u8));
+
+        d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                          vreinterpret_u16_u32(d2tmp2.val[0]));
+        d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                          vreinterpret_u16_u32(d2tmp3.val[0]));
+        d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                          vreinterpret_u16_u32(d2tmp2.val[1]));
+        d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                          vreinterpret_u16_u32(d2tmp3.val[1]));
+
+        d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                         vreinterpret_u8_u16(d2tmp5.val[0]));
+        d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                         vreinterpret_u8_u16(d2tmp5.val[1]));
+        d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                          vreinterpret_u8_u16(d2tmp7.val[0]));
+        d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                          vreinterpret_u8_u16(d2tmp7.val[1]));
+
+        d3u8 = d2tmp8.val[0];
+        d4u8 = d2tmp8.val[1];
+        d5u8 = d2tmp9.val[0];
+        d6u8 = d2tmp9.val[1];
+        d7u8 = d2tmp10.val[0];
+        d16u8 = d2tmp10.val[1];
+        d17u8 = d2tmp11.val[0];
+        d18u8 = d2tmp11.val[1];
+
+        vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
+                             d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+                             &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+        d4Result.val[0] = d0u8;
+        d4Result.val[1] = d1u8;
+        d4Result.val[2] = d2u8;
+        d4Result.val[3] = d3u8;
+
+        d2Result.val[0] = d4u8;
+        d2Result.val[1] = d5u8;
+
+        s = src - 3;
+        vst4_lane_u8(s, d4Result, 0);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 1);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 2);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 3);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 4);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 5);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 6);
+        s += pitch;
+        vst4_lane_u8(s, d4Result, 7);
+
+        s = src + 1;
+        vst2_lane_u8(s, d2Result, 0);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 1);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 2);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 3);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 4);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 5);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 6);
+        s += pitch;
+        vst2_lane_u8(s, d2Result, 7);
+    }
+    return;
+}
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon_asm.asm
new file mode 100644
index 000000000..91aaec04e
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_8_neon_asm.asm
@@ -0,0 +1,451 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_lpf_horizontal_8_neon|
+    EXPORT  |vp9_lpf_vertical_8_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_horizontal_8_neon(uint8_t *s, int p,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh,
+;                                int count)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; sp+4  int count
+|vp9_lpf_horizontal_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r12, [sp, #16]             ; load count
+    ldr         r2, [sp, #12]              ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    cmp         r12, #0
+    beq         end_vp9_mblf_h_edge
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+count_mblf_h_loop
+    sub         r3, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r2, r3, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r3@64], r1          ; p3
+    vld1.u8     {d4}, [r2@64], r1          ; p2
+    vld1.u8     {d5}, [r3@64], r1          ; p1
+    vld1.u8     {d6}, [r2@64], r1          ; p0
+    vld1.u8     {d7}, [r3@64], r1          ; q0
+    vld1.u8     {d16}, [r2@64], r1         ; q1
+    vld1.u8     {d17}, [r3@64]             ; q2
+    vld1.u8     {d18}, [r2@64], r1         ; q3
+
+    sub         r3, r3, r1, lsl #1
+    sub         r2, r2, r1, lsl #2
+
+    bl          vp9_mbloop_filter_neon
+
+    vst1.u8     {d0}, [r2@64], r1          ; store op2
+    vst1.u8     {d1}, [r3@64], r1          ; store op1
+    vst1.u8     {d2}, [r2@64], r1          ; store op0
+    vst1.u8     {d3}, [r3@64], r1          ; store oq0
+    vst1.u8     {d4}, [r2@64], r1          ; store oq1
+    vst1.u8     {d5}, [r3@64], r1          ; store oq2
+
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         count_mblf_h_loop
+
+end_vp9_mblf_h_edge
+    pop         {r4-r5, pc}
+
+    ENDP        ; |vp9_lpf_horizontal_8_neon|
+
+; void vp9_lpf_vertical_8_neon(uint8_t *s,
+;                              int pitch,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh,
+;                              int count)
+;
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; sp+4  int count
+|vp9_lpf_vertical_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    ldr         r12, [sp, #16]            ; load count
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #12]             ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+    cmp         r12, #0
+    beq         end_vp9_mblf_v_edge
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+count_mblf_v_loop
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    sub         r2, r0, #3
+    add         r3, r0, #1
+
+    bl          vp9_mbloop_filter_neon
+
+    ;store op2, op1, op0, oq0
+    vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r2], r1
+    vst4.8      {d0[1], d1[1], d2[1], d3[1]}, [r2], r1
+    vst4.8      {d0[2], d1[2], d2[2], d3[2]}, [r2], r1
+    vst4.8      {d0[3], d1[3], d2[3], d3[3]}, [r2], r1
+    vst4.8      {d0[4], d1[4], d2[4], d3[4]}, [r2], r1
+    vst4.8      {d0[5], d1[5], d2[5], d3[5]}, [r2], r1
+    vst4.8      {d0[6], d1[6], d2[6], d3[6]}, [r2], r1
+    vst4.8      {d0[7], d1[7], d2[7], d3[7]}, [r2]
+
+    ;store oq1, oq2
+    vst2.8      {d4[0], d5[0]}, [r3], r1
+    vst2.8      {d4[1], d5[1]}, [r3], r1
+    vst2.8      {d4[2], d5[2]}, [r3], r1
+    vst2.8      {d4[3], d5[3]}, [r3], r1
+    vst2.8      {d4[4], d5[4]}, [r3], r1
+    vst2.8      {d4[5], d5[5]}, [r3], r1
+    vst2.8      {d4[6], d5[6]}, [r3], r1
+    vst2.8      {d4[7], d5[7]}, [r3]
+
+    add         r0, r0, r1, lsl #3         ; s += pitch * 8
+    subs        r12, r12, #1
+    subne       r2, r0, #4                 ; move s pointer down by 4 columns
+    bne         count_mblf_v_loop
+
+end_vp9_mblf_v_edge
+    pop         {r4-r5, pc}
+    ENDP        ; |vp9_lpf_vertical_8_neon|
+
+; void vp9_mbloop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d0    op2
+; d1    op1
+; d2    op0
+; d3    oq0
+; d4    oq1
+; d5    oq2
+|vp9_mbloop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7               ; m4 = abs(q1 - q0)
+    vabd.u8     d23, d17, d16              ; m5 = abs(q2 - q1)
+    vabd.u8     d24, d18, d17              ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22              ; m2 = max(m3, m4)
+
+    vabd.u8     d25, d6, d4                ; m7 = abs(p0 - p2)
+
+    vmax.u8     d23, d23, d24              ; m3 = max(m5, m6)
+
+    vabd.u8     d26, d7, d17               ; m8 = abs(q0 - q2)
+
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d6, d7                ; m9 = abs(p0 - q0)
+    vabd.u8     d27, d3, d6                ; m10 = abs(p3 - p0)
+    vabd.u8     d28, d18, d7               ; m11 = abs(q3 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d5, d16               ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d1, d19
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; m4 = max(m7, m8)
+    vmax.u8     d26, d27, d28              ; m5 = max(m10, m11)
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+
+    vmax.u8     d25, d25, d26              ; m4 = max(m4, m5)
+
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmax.u8     d20, d20, d25              ; m2 = max(m2, m4)
+
+    vmov.u8     d23, #1
+    vcge.u8     d24, d0, d24               ; a > blimit
+
+    vcgt.u8     d21, d21, d2               ; (abs(p1 - p0) > thresh)*-1
+
+    vcge.u8     d20, d23, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    vcgt.u8     d23, d22, d2               ; (abs(q1 - q0) > thresh)*-1
+
+    vand        d20, d20, d19              ; flat & mask
+
+    vmov.u8     d22, #0x80
+
+    vorr        d23, d21, d23              ; hev
+
+    ; This instruction will truncate the "flat & mask" masks down to 4 bits
+    ; each to fit into one 32 bit arm register. The values are stored in
+    ; q10.64[0].
+    vshrn.u16   d30, q10, #4
+    vmov.u32    r4, d30[0]                 ; flat & mask 4bits
+
+    adds        r5, r4, #1                 ; Check for all 1's
+
+    ; If mask and flat are 1's for all vectors, then we only need to execute
+    ; the power branch for all vectors.
+    beq         power_branch_only
+
+    cmp         r4, #0                     ; Check for 0, set flag for later
+
+    ; mbfilter() function
+    ; filter() function
+    ; convert to signed
+    veor        d21, d7, d22               ; qs0
+    veor        d24, d6, d22               ; ps0
+    veor        d25, d5, d22               ; ps1
+    veor        d26, d16, d22              ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d21, d24              ; ( qs0 - ps0)
+
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+
+    vand        d29, d29, d23              ; filter &= hev
+
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d21, d21, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d23              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    ; If mask and flat are 0's for all vectors, then we only need to execute
+    ; the filter branch for all vectors.
+    beq         filter_branch_only
+
+    ; If mask and flat are mixed then we must perform both branches and
+    ; combine the data.
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d21, d21, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    ; At this point we have already executed the filter branch. The filter
+    ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
+    ; branch and combine the data.
+    vmov.u8     d23, #2
+    vaddl.u8    q14, d6, d7                ; r_op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; r_op2 += p3 * 3
+    vmlal.u8    q14, d4, d23               ; r_op2 += p2 * 2
+
+    vbif        d0, d4, d20                ; op2 |= p2 & ~(flat & mask)
+
+    vaddw.u8    q14, d5                    ; r_op2 += p1
+
+    vbif        d1, d25, d20               ; op1 |= f_op1 & ~(flat & mask)
+
+    vqrshrn.u16 d30, q14, #3               ; r_op2
+
+    vsubw.u8    q14, d3                    ; r_op1 = r_op2 - p3
+    vsubw.u8    q14, d4                    ; r_op1 -= p2
+    vaddw.u8    q14, d5                    ; r_op1 += p1
+    vaddw.u8    q14, d16                   ; r_op1 += q1
+
+    vbif        d2, d24, d20               ; op0 |= f_op0 & ~(flat & mask)
+
+    vqrshrn.u16 d31, q14, #3               ; r_op1
+
+    vsubw.u8    q14, d3                    ; r_op0 = r_op1 - p3
+    vsubw.u8    q14, d5                    ; r_op0 -= p1
+    vaddw.u8    q14, d6                    ; r_op0 += p0
+    vaddw.u8    q14, d17                   ; r_op0 += q2
+
+    vbit        d0, d30, d20               ; op2 |= r_op2 & (flat & mask)
+
+    vqrshrn.u16 d23, q14, #3               ; r_op0
+
+    vsubw.u8    q14, d3                    ; r_oq0 = r_op0 - p3
+    vsubw.u8    q14, d6                    ; r_oq0 -= p0
+    vaddw.u8    q14, d7                    ; r_oq0 += q0
+
+    vbit        d1, d31, d20               ; op1 |= r_op1 & (flat & mask)
+
+    vaddw.u8    q14, d18                   ; oq0 += q3
+
+    vbit        d2, d23, d20               ; op0 |= r_op0 & (flat & mask)
+
+    vqrshrn.u16 d22, q14, #3               ; r_oq0
+
+    vsubw.u8    q14, d4                    ; r_oq1 = r_oq0 - p2
+    vsubw.u8    q14, d7                    ; r_oq1 -= q0
+    vaddw.u8    q14, d16                   ; r_oq1 += q1
+
+    vbif        d3, d21, d20               ; oq0 |= f_oq0 & ~(flat & mask)
+
+    vaddw.u8    q14, d18                   ; r_oq1 += q3
+
+    vbif        d4, d26, d20               ; oq1 |= f_oq1 & ~(flat & mask)
+
+    vqrshrn.u16 d6, q14, #3                ; r_oq1
+
+    vsubw.u8    q14, d5                    ; r_oq2 = r_oq1 - p1
+    vsubw.u8    q14, d16                   ; r_oq2 -= q1
+    vaddw.u8    q14, d17                   ; r_oq2 += q2
+    vaddw.u8    q14, d18                   ; r_oq2 += q3
+
+    vbif        d5, d17, d20               ; oq2 |= q2 & ~(flat & mask)
+
+    vqrshrn.u16 d7, q14, #3                ; r_oq2
+
+    vbit        d3, d22, d20               ; oq0 |= r_oq0 & (flat & mask)
+    vbit        d4, d6, d20                ; oq1 |= r_oq1 & (flat & mask)
+    vbit        d5, d7, d20                ; oq2 |= r_oq2 & (flat & mask)
+
+    bx          lr
+
+power_branch_only
+    vmov.u8     d27, #3
+    vmov.u8     d21, #2
+    vaddl.u8    q14, d6, d7                ; op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; op2 += p3 * 3
+    vmlal.u8    q14, d4, d21               ; op2 += p2 * 2
+    vaddw.u8    q14, d5                    ; op2 += p1
+    vqrshrn.u16 d0, q14, #3                ; op2
+
+    vsubw.u8    q14, d3                    ; op1 = op2 - p3
+    vsubw.u8    q14, d4                    ; op1 -= p2
+    vaddw.u8    q14, d5                    ; op1 += p1
+    vaddw.u8    q14, d16                   ; op1 += q1
+    vqrshrn.u16 d1, q14, #3                ; op1
+
+    vsubw.u8    q14, d3                    ; op0 = op1 - p3
+    vsubw.u8    q14, d5                    ; op0 -= p1
+    vaddw.u8    q14, d6                    ; op0 += p0
+    vaddw.u8    q14, d17                   ; op0 += q2
+    vqrshrn.u16 d2, q14, #3                ; op0
+
+    vsubw.u8    q14, d3                    ; oq0 = op0 - p3
+    vsubw.u8    q14, d6                    ; oq0 -= p0
+    vaddw.u8    q14, d7                    ; oq0 += q0
+    vaddw.u8    q14, d18                   ; oq0 += q3
+    vqrshrn.u16 d3, q14, #3                ; oq0
+
+    vsubw.u8    q14, d4                    ; oq1 = oq0 - p2
+    vsubw.u8    q14, d7                    ; oq1 -= q0
+    vaddw.u8    q14, d16                   ; oq1 += q1
+    vaddw.u8    q14, d18                   ; oq1 += q3
+    vqrshrn.u16 d4, q14, #3                ; oq1
+
+    vsubw.u8    q14, d5                    ; oq2 = oq1 - p1
+    vsubw.u8    q14, d16                   ; oq2 -= q1
+    vaddw.u8    q14, d17                   ; oq2 += q2
+    vaddw.u8    q14, d18                   ; oq2 += q3
+    vqrshrn.u16 d5, q14, #3                ; oq2
+
+    bx          lr
+
+filter_branch_only
+    ; TODO(fgalligan): See if we can rearange registers so we do not need to
+    ; do the 2 vswp.
+    vswp        d0, d4                      ; op2
+    vswp        d5, d17                     ; oq2
+    veor        d2, d24, d22                ; *op0 = u^0x80
+    veor        d3, d21, d22                ; *oq0 = u^0x80
+    veor        d1, d25, d22                ; *op1 = u^0x80
+    veor        d4, d26, d22                ; *oq1 = u^0x80
+
+    bx          lr
+
+    ENDP        ; |vp9_mbloop_filter_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.c
new file mode 100644
index 000000000..31fcc63ba
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_loopfilter_neon.c
@@ -0,0 +1,58 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+void vp9_lpf_vertical_4_dual_neon(uint8_t *s, int p,
+                                  const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  vp9_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0, 1);
+  vp9_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
+}
+
+#if HAVE_NEON_ASM
+void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
+                                    const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+  vp9_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0, 1);
+  vp9_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_8_dual_neon(uint8_t *s, int p,
+                                  const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  vp9_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0, 1);
+  vp9_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_16_dual_neon(uint8_t *s, int p,
+                                   const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  vp9_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
+  vp9_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // HAVE_NEON_ASM
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm b/media/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
new file mode 100644
index 000000000..5fe2bba46
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
@@ -0,0 +1,606 @@
+;
+;  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_lpf_horizontal_16_neon|
+    EXPORT  |vp9_lpf_vertical_16_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; void vp9_lpf_horizontal_16_neon(uint8_t *s, int p,
+;                                 const uint8_t *blimit,
+;                                 const uint8_t *limit,
+;                                 const uint8_t *thresh
+;                                 int count)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vp9_lpf_horizontal_16_neon| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+    ldr         r12, [sp, #92]             ; load count
+
+h_count
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, r1, lsl #3         ; move src pointer down by 8 lines
+
+    vld1.u8     {d0}, [r8@64], r1          ; p7
+    vld1.u8     {d1}, [r8@64], r1          ; p6
+    vld1.u8     {d2}, [r8@64], r1          ; p5
+    vld1.u8     {d3}, [r8@64], r1          ; p4
+    vld1.u8     {d4}, [r8@64], r1          ; p3
+    vld1.u8     {d5}, [r8@64], r1          ; p2
+    vld1.u8     {d6}, [r8@64], r1          ; p1
+    vld1.u8     {d7}, [r8@64], r1          ; p0
+    vld1.u8     {d8}, [r8@64], r1          ; q0
+    vld1.u8     {d9}, [r8@64], r1          ; q1
+    vld1.u8     {d10}, [r8@64], r1         ; q2
+    vld1.u8     {d11}, [r8@64], r1         ; q3
+    vld1.u8     {d12}, [r8@64], r1         ; q4
+    vld1.u8     {d13}, [r8@64], r1         ; q5
+    vld1.u8     {d14}, [r8@64], r1         ; q6
+    vld1.u8     {d15}, [r8@64], r1         ; q7
+
+    bl          vp9_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         h_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, r1, lsl #1
+
+    vst1.u8     {d25}, [r8@64], r1         ; store op1
+    vst1.u8     {d24}, [r8@64], r1         ; store op0
+    vst1.u8     {d23}, [r8@64], r1         ; store oq0
+    vst1.u8     {d26}, [r8@64], r1         ; store oq1
+
+    b           h_next
+
+h_mbfilter
+    tst         r7, #2
+    beq         h_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, r1, lsl #1
+    sub         r8, r8, r1
+
+    vst1.u8     {d18}, [r8@64], r1         ; store op2
+    vst1.u8     {d19}, [r8@64], r1         ; store op1
+    vst1.u8     {d20}, [r8@64], r1         ; store op0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq0
+    vst1.u8     {d22}, [r8@64], r1         ; store oq1
+    vst1.u8     {d23}, [r8@64], r1         ; store oq2
+
+    b           h_next
+
+h_wide_mbfilter
+    sub         r8, r0, r1, lsl #3
+    add         r8, r8, r1
+
+    vst1.u8     {d16}, [r8@64], r1         ; store op6
+    vst1.u8     {d24}, [r8@64], r1         ; store op5
+    vst1.u8     {d25}, [r8@64], r1         ; store op4
+    vst1.u8     {d26}, [r8@64], r1         ; store op3
+    vst1.u8     {d27}, [r8@64], r1         ; store op2
+    vst1.u8     {d18}, [r8@64], r1         ; store op1
+    vst1.u8     {d19}, [r8@64], r1         ; store op0
+    vst1.u8     {d20}, [r8@64], r1         ; store oq0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq1
+    vst1.u8     {d22}, [r8@64], r1         ; store oq2
+    vst1.u8     {d23}, [r8@64], r1         ; store oq3
+    vst1.u8     {d1}, [r8@64], r1          ; store oq4
+    vst1.u8     {d2}, [r8@64], r1          ; store oq5
+    vst1.u8     {d3}, [r8@64], r1          ; store oq6
+
+h_next
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         h_count
+
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |vp9_lpf_horizontal_16_neon|
+
+; void vp9_lpf_vertical_16_neon(uint8_t *s, int p,
+;                               const uint8_t *blimit,
+;                               const uint8_t *limit,
+;                               const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|vp9_lpf_vertical_16_neon| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, #8
+
+    vld1.8      {d0}, [r8@64], r1
+    vld1.8      {d8}, [r0@64], r1
+    vld1.8      {d1}, [r8@64], r1
+    vld1.8      {d9}, [r0@64], r1
+    vld1.8      {d2}, [r8@64], r1
+    vld1.8      {d10}, [r0@64], r1
+    vld1.8      {d3}, [r8@64], r1
+    vld1.8      {d11}, [r0@64], r1
+    vld1.8      {d4}, [r8@64], r1
+    vld1.8      {d12}, [r0@64], r1
+    vld1.8      {d5}, [r8@64], r1
+    vld1.8      {d13}, [r0@64], r1
+    vld1.8      {d6}, [r8@64], r1
+    vld1.8      {d14}, [r0@64], r1
+    vld1.8      {d7}, [r8@64], r1
+    vld1.8      {d15}, [r0@64], r1
+
+    sub         r0, r0, r1, lsl #3
+
+    vtrn.32     q0, q2
+    vtrn.32     q1, q3
+    vtrn.32     q4, q6
+    vtrn.32     q5, q7
+
+    vtrn.16     q0, q1
+    vtrn.16     q2, q3
+    vtrn.16     q4, q5
+    vtrn.16     q6, q7
+
+    vtrn.8      d0, d1
+    vtrn.8      d2, d3
+    vtrn.8      d4, d5
+    vtrn.8      d6, d7
+
+    vtrn.8      d8, d9
+    vtrn.8      d10, d11
+    vtrn.8      d12, d13
+    vtrn.8      d14, d15
+
+    bl          vp9_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         v_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, #2
+
+    vswp        d23, d25
+
+    vst4.8      {d23[0], d24[0], d25[0], d26[0]}, [r8], r1
+    vst4.8      {d23[1], d24[1], d25[1], d26[1]}, [r8], r1
+    vst4.8      {d23[2], d24[2], d25[2], d26[2]}, [r8], r1
+    vst4.8      {d23[3], d24[3], d25[3], d26[3]}, [r8], r1
+    vst4.8      {d23[4], d24[4], d25[4], d26[4]}, [r8], r1
+    vst4.8      {d23[5], d24[5], d25[5], d26[5]}, [r8], r1
+    vst4.8      {d23[6], d24[6], d25[6], d26[6]}, [r8], r1
+    vst4.8      {d23[7], d24[7], d25[7], d26[7]}, [r8], r1
+
+    b           v_end
+
+v_mbfilter
+    tst         r7, #2
+    beq         v_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, #3
+
+    vst3.8      {d18[0], d19[0], d20[0]}, [r8], r1
+    vst3.8      {d21[0], d22[0], d23[0]}, [r0], r1
+    vst3.8      {d18[1], d19[1], d20[1]}, [r8], r1
+    vst3.8      {d21[1], d22[1], d23[1]}, [r0], r1
+    vst3.8      {d18[2], d19[2], d20[2]}, [r8], r1
+    vst3.8      {d21[2], d22[2], d23[2]}, [r0], r1
+    vst3.8      {d18[3], d19[3], d20[3]}, [r8], r1
+    vst3.8      {d21[3], d22[3], d23[3]}, [r0], r1
+    vst3.8      {d18[4], d19[4], d20[4]}, [r8], r1
+    vst3.8      {d21[4], d22[4], d23[4]}, [r0], r1
+    vst3.8      {d18[5], d19[5], d20[5]}, [r8], r1
+    vst3.8      {d21[5], d22[5], d23[5]}, [r0], r1
+    vst3.8      {d18[6], d19[6], d20[6]}, [r8], r1
+    vst3.8      {d21[6], d22[6], d23[6]}, [r0], r1
+    vst3.8      {d18[7], d19[7], d20[7]}, [r8], r1
+    vst3.8      {d21[7], d22[7], d23[7]}, [r0], r1
+
+    b           v_end
+
+v_wide_mbfilter
+    sub         r8, r0, #8
+
+    vtrn.32     d0,  d26
+    vtrn.32     d16, d27
+    vtrn.32     d24, d18
+    vtrn.32     d25, d19
+
+    vtrn.16     d0,  d24
+    vtrn.16     d16, d25
+    vtrn.16     d26, d18
+    vtrn.16     d27, d19
+
+    vtrn.8      d0,  d16
+    vtrn.8      d24, d25
+    vtrn.8      d26, d27
+    vtrn.8      d18, d19
+
+    vtrn.32     d20, d1
+    vtrn.32     d21, d2
+    vtrn.32     d22, d3
+    vtrn.32     d23, d15
+
+    vtrn.16     d20, d22
+    vtrn.16     d21, d23
+    vtrn.16     d1,  d3
+    vtrn.16     d2,  d15
+
+    vtrn.8      d20, d21
+    vtrn.8      d22, d23
+    vtrn.8      d1,  d2
+    vtrn.8      d3,  d15
+
+    vst1.8      {d0}, [r8@64], r1
+    vst1.8      {d20}, [r0@64], r1
+    vst1.8      {d16}, [r8@64], r1
+    vst1.8      {d21}, [r0@64], r1
+    vst1.8      {d24}, [r8@64], r1
+    vst1.8      {d22}, [r0@64], r1
+    vst1.8      {d25}, [r8@64], r1
+    vst1.8      {d23}, [r0@64], r1
+    vst1.8      {d26}, [r8@64], r1
+    vst1.8      {d1}, [r0@64], r1
+    vst1.8      {d27}, [r8@64], r1
+    vst1.8      {d2}, [r0@64], r1
+    vst1.8      {d18}, [r8@64], r1
+    vst1.8      {d3}, [r0@64], r1
+    vst1.8      {d19}, [r8@64], r1
+    vst1.8      {d15}, [r0@64], r1
+
+v_end
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |vp9_lpf_vertical_16_neon|
+
+; void vp9_wide_mbfilter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store.
+;
+; r0-r3 PRESERVE
+; d16    blimit
+; d17    limit
+; d18    thresh
+; d0    p7
+; d1    p6
+; d2    p5
+; d3    p4
+; d4    p3
+; d5    p2
+; d6    p1
+; d7    p0
+; d8    q0
+; d9    q1
+; d10   q2
+; d11   q3
+; d12   q4
+; d13   q5
+; d14   q6
+; d15   q7
+|vp9_wide_mbfilter_neon| PROC
+    mov         r7, #0
+
+    ; filter_mask
+    vabd.u8     d19, d4, d5                ; abs(p3 - p2)
+    vabd.u8     d20, d5, d6                ; abs(p2 - p1)
+    vabd.u8     d21, d6, d7                ; abs(p1 - p0)
+    vabd.u8     d22, d9, d8                ; abs(q1 - q0)
+    vabd.u8     d23, d10, d9               ; abs(q2 - q1)
+    vabd.u8     d24, d11, d10              ; abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; max(abs(p3 - p2), abs(p2 - p1))
+    vmax.u8     d20, d21, d22              ; max(abs(p1 - p0), abs(q1 - q0))
+    vmax.u8     d23, d23, d24              ; max(abs(q2 - q1), abs(q3 - q2))
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d7, d8                ; abs(p0 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d6, d9                ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d17, d19
+
+    ; flatmask4
+    vabd.u8     d25, d7, d5                ; abs(p0 - p2)
+    vabd.u8     d26, d8, d10               ; abs(q0 - q2)
+    vabd.u8     d27, d4, d7                ; abs(p3 - p0)
+    vabd.u8     d28, d11, d8               ; abs(q3 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; max(abs(p0 - p2), abs(q0 - q2))
+    vmax.u8     d26, d27, d28              ; max(abs(p3 - p0), abs(q3 - q0))
+    vmax.u8     d25, d25, d26
+    vmax.u8     d20, d20, d25
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmov.u8     d30, #1
+    vcge.u8     d24, d16, d24              ; (a > blimit * 2 + limit) * -1
+
+    vcge.u8     d20, d30, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    ; hevmask
+    vcgt.u8     d21, d21, d18              ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d18              ; (abs(q1 - q0) > thresh)*-1
+    vorr        d21, d21, d22              ; hev
+
+    vand        d16, d20, d19              ; flat && mask
+    vmov        r5, r6, d16
+
+    ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
+    vabd.u8     d22, d3, d7                ; abs(p4 - p0)
+    vabd.u8     d23, d12, d8               ; abs(q4 - q0)
+    vabd.u8     d24, d7, d2                ; abs(p0 - p5)
+    vabd.u8     d25, d8, d13               ; abs(q0 - q5)
+    vabd.u8     d26, d1, d7                ; abs(p6 - p0)
+    vabd.u8     d27, d14, d8               ; abs(q6 - q0)
+    vabd.u8     d28, d0, d7                ; abs(p7 - p0)
+    vabd.u8     d29, d15, d8               ; abs(q7 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d22, d22, d23              ; max(abs(p4 - p0), abs(q4 - q0))
+    vmax.u8     d23, d24, d25              ; max(abs(p0 - p5), abs(q0 - q5))
+    vmax.u8     d24, d26, d27              ; max(abs(p6 - p0), abs(q6 - q0))
+    vmax.u8     d25, d28, d29              ; max(abs(p7 - p0), abs(q7 - q0))
+
+    vmax.u8     d26, d22, d23
+    vmax.u8     d27, d24, d25
+    vmax.u8     d23, d26, d27
+
+    vcge.u8     d18, d30, d23              ; flat2
+
+    vmov.u8     d22, #0x80
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #1                 ; Only do filter branch
+
+    vand        d17, d18, d16              ; flat2 && flat && mask
+    vmov        r5, r6, d17
+
+    ; mbfilter() function
+
+    ; filter() function
+    ; convert to signed
+    veor        d23, d8, d22               ; qs0
+    veor        d24, d7, d22               ; ps0
+    veor        d25, d6, d22               ; ps1
+    veor        d26, d9, d22               ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d23, d24              ; ( qs0 - ps0)
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+    vand        d29, d29, d21              ; filter &= hev
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d23, d23, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d21              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d23, d23, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    tst         r7, #1
+    bxne        lr
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #2                 ; Only do mbfilter branch
+
+    ; mbfilter flat && mask branch
+    ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
+    ; and using vibt on the q's?
+    vmov.u8     d29, #2
+    vaddl.u8    q15, d7, d8                ; op2 = p0 + q0
+    vmlal.u8    q15, d4, d27               ; op2 = p0 + q0 + p3 * 3
+    vmlal.u8    q15, d5, d29               ; op2 = p0 + q0 + p3 * 3 + p2 * 2
+    vaddl.u8    q10, d4, d5
+    vaddw.u8    q15, d6                    ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
+    vaddl.u8    q14, d6, d9
+    vqrshrn.u16 d18, q15, #3               ; r_op2
+
+    vsub.i16    q15, q10
+    vaddl.u8    q10, d4, d6
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d7, d10
+    vqrshrn.u16 d19, q15, #3               ; r_op1
+
+    vsub.i16    q15, q10
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d8, d11
+    vqrshrn.u16 d20, q15, #3               ; r_op0
+
+    vsubw.u8    q15, d4                    ; oq0 = op0 - p3
+    vsubw.u8    q15, d7                    ; oq0 -= p0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d9, d11
+    vqrshrn.u16 d21, q15, #3               ; r_oq0
+
+    vsubw.u8    q15, d5                    ; oq1 = oq0 - p2
+    vsubw.u8    q15, d8                    ; oq1 -= q0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d10, d11
+    vqrshrn.u16 d22, q15, #3               ; r_oq1
+
+    vsubw.u8    q15, d6                    ; oq2 = oq0 - p1
+    vsubw.u8    q15, d9                    ; oq2 -= q1
+    vadd.i16    q15, q14
+    vqrshrn.u16 d27, q15, #3               ; r_oq2
+
+    ; Filter does not set op2 or oq2, so use p2 and q2.
+    vbif        d18, d5, d16               ; t_op2 |= p2 & ~(flat & mask)
+    vbif        d19, d25, d16              ; t_op1 |= f_op1 & ~(flat & mask)
+    vbif        d20, d24, d16              ; t_op0 |= f_op0 & ~(flat & mask)
+    vbif        d21, d23, d16              ; t_oq0 |= f_oq0 & ~(flat & mask)
+    vbif        d22, d26, d16              ; t_oq1 |= f_oq1 & ~(flat & mask)
+
+    vbit        d23, d27, d16              ; t_oq2 |= r_oq2 & (flat & mask)
+    vbif        d23, d10, d16              ; t_oq2 |= q2 & ~(flat & mask)
+
+    tst         r7, #2
+    bxne        lr
+
+    ; wide_mbfilter flat2 && flat && mask branch
+    vmov.u8     d16, #7
+    vaddl.u8    q15, d7, d8                ; op6 = p0 + q0
+    vaddl.u8    q12, d2, d3
+    vaddl.u8    q13, d4, d5
+    vaddl.u8    q14, d1, d6
+    vmlal.u8    q15, d0, d16               ; op6 += p7 * 3
+    vadd.i16    q12, q13
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vadd.i16    q15, q12
+    vaddl.u8    q12, d0, d1
+    vaddw.u8    q15, d1
+    vaddl.u8    q13, d0, d2
+    vadd.i16    q14, q15, q14
+    vqrshrn.u16 d16, q15, #4               ; w_op6
+
+    vsub.i16    q15, q14, q12
+    vaddl.u8    q14, d3, d10
+    vqrshrn.u16 d24, q15, #4               ; w_op5
+
+    vsub.i16    q15, q13
+    vaddl.u8    q13, d0, d3
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vqrshrn.u16 d25, q15, #4               ; w_op4
+
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d0, d4
+    vsub.i16    q15, q13
+    vsub.i16    q14, q15, q14
+    vqrshrn.u16 d26, q15, #4               ; w_op3
+
+    vaddw.u8    q15, q14, d5               ; op2 += p2
+    vaddl.u8    q14, d0, d5
+    vaddw.u8    q15, d12                   ; op2 += q4
+    vbif        d26, d4, d17               ; op3 |= p3 & ~(f2 & f & m)
+    vqrshrn.u16 d27, q15, #4               ; w_op2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d6
+    vaddw.u8    q15, d6                    ; op1 += p1
+    vaddw.u8    q15, d13                   ; op1 += q5
+    vbif        d27, d18, d17              ; op2 |= t_op2 & ~(f2 & f & m)
+    vqrshrn.u16 d18, q15, #4               ; w_op1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d7
+    vaddw.u8    q15, d7                    ; op0 += p0
+    vaddw.u8    q15, d14                   ; op0 += q6
+    vbif        d18, d19, d17              ; op1 |= t_op1 & ~(f2 & f & m)
+    vqrshrn.u16 d19, q15, #4               ; w_op0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d1, d8
+    vaddw.u8    q15, d8                    ; oq0 += q0
+    vaddw.u8    q15, d15                   ; oq0 += q7
+    vbif        d19, d20, d17              ; op0 |= t_op0 & ~(f2 & f & m)
+    vqrshrn.u16 d20, q15, #4               ; w_oq0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vaddw.u8    q15, d9                    ; oq1 += q1
+    vaddl.u8    q4, d10, d15
+    vaddw.u8    q15, d15                   ; oq1 += q7
+    vbif        d20, d21, d17              ; oq0 |= t_oq0 & ~(f2 & f & m)
+    vqrshrn.u16 d21, q15, #4               ; w_oq1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d3, d10
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d11, d15
+    vbif        d21, d22, d17              ; oq1 |= t_oq1 & ~(f2 & f & m)
+    vqrshrn.u16 d22, q15, #4               ; w_oq2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d12, d15
+    vbif        d22, d23, d17              ; oq2 |= t_oq2 & ~(f2 & f & m)
+    vqrshrn.u16 d23, q15, #4               ; w_oq3
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d5, d12
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d13, d15
+    vbif        d16, d1, d17               ; op6 |= p6 & ~(f2 & f & m)
+    vqrshrn.u16 d1, q15, #4                ; w_oq4
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d6, d13
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d14, d15
+    vbif        d24, d2, d17               ; op5 |= p5 & ~(f2 & f & m)
+    vqrshrn.u16 d2, q15, #4                ; w_oq5
+
+    vsub.i16    q15, q14
+    vbif        d25, d3, d17               ; op4 |= p4 & ~(f2 & f & m)
+    vadd.i16    q15, q4
+    vbif        d23, d11, d17              ; oq3 |= q3 & ~(f2 & f & m)
+    vqrshrn.u16 d3, q15, #4                ; w_oq6
+    vbif        d1, d12, d17               ; oq4 |= q4 & ~(f2 & f & m)
+    vbif        d2, d13, d17               ; oq5 |= q5 & ~(f2 & f & m)
+    vbif        d3, d14, d17               ; oq6 |= q6 & ~(f2 & f & m)
+
+    bx          lr
+    ENDP        ; |vp9_wide_mbfilter_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.c b/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.c
new file mode 100644
index 000000000..499c42ac3
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon.c
@@ -0,0 +1,578 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+//------------------------------------------------------------------------------
+// DC 8x8
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride,
+                          const uint8_t *above, const uint8_t *left,
+                          int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_top = vcombine_u16(p2, p2);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);  // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_left = vcombine_u16(p2, p2);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 4);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 3);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 3);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 8; ++i) {
+      vst1_u32((uint32_t*)(dst + i * stride), vreinterpret_u32_u8(dc));
+    }
+  }
+}
+
+void vp9_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_8x8(dst, stride, above, left, 1, 1);
+}
+
+void vp9_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_8x8(dst, stride, NULL, left, 0, 1);
+}
+
+void vp9_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_8x8(dst, stride, above, NULL, 1, 0);
+}
+
+void vp9_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_8x8(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 16x16
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A = vld1q_u8(above);  // top row
+    const uint16x8_t p0 = vpaddlq_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_top = vcombine_u16(p3, p3);
+  }
+
+  if (do_left) {
+    const uint8x16_t L = vld1q_u8(left);  // left row
+    const uint16x8_t p0 = vpaddlq_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_left = vcombine_u16(p3, p3);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 5);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 4);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 4);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 16; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+    }
+  }
+}
+
+void vp9_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_16x16(dst, stride, above, left, 1, 1);
+}
+
+void vp9_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_16x16(dst, stride, NULL, left, 0, 1);
+}
+
+void vp9_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_16x16(dst, stride, above, NULL, 1, 0);
+}
+
+void vp9_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_16x16(dst, stride, NULL, NULL, 0, 0);
+}
+
+#if !HAVE_NEON_ASM
+
+void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint32x2_t d0u32 = vdup_n_u32(0);
+  (void)left;
+
+  d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
+  for (i = 0; i < 4; i++, dst += stride)
+    vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+}
+
+void vp9_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  (void)left;
+
+  d0u8 = vld1_u8(above);
+  for (i = 0; i < 8; i++, dst += stride)
+    vst1_u8(dst, d0u8);
+}
+
+void vp9_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  for (i = 0; i < 16; i++, dst += stride)
+    vst1q_u8(dst, q0u8);
+}
+
+void vp9_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  q1u8 = vld1q_u8(above + 16);
+  for (i = 0; i < 32; i++, dst += stride) {
+    vst1q_u8(dst, q0u8);
+    vst1q_u8(dst + 16, q1u8);
+  }
+}
+
+void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d1u32 = vdup_n_u32(0);
+  (void)above;
+
+  d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+}
+
+void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint64x1_t d1u64 = vdup_n_u64(0);
+  (void)above;
+
+  d1u64 = vld1_u64((const uint64_t *)left);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
+  vst1_u8(dst, d0u8);
+}
+
+void vp9_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  q1u8 = vld1q_u8(left);
+  d2u8 = vget_low_u8(q1u8);
+  for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+    q0u8 = vdupq_lane_u8(d2u8, 0);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 1);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 2);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 3);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 4);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 5);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 6);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 7);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+  }
+}
+
+void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  for (k = 0; k < 2; k++, left += 16) {
+    q1u8 = vld1q_u8(left);
+    d2u8 = vget_low_u8(q1u8);
+    for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+      q0u8 = vdupq_lane_u8(d2u8, 0);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 1);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 2);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 3);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 4);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 5);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 6);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 7);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+    }
+  }
+}
+
+void vp9_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint16x8_t q1u16, q3u16;
+  int16x8_t q1s16;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d2u32 = vdup_n_u32(0);
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
+  q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
+  for (i = 0; i < 4; i++, dst += stride) {
+    q1u16 = vdupq_n_u16((uint16_t)left[i]);
+    q1s16 = vaddq_s16(vreinterpretq_s16_u16(q1u16),
+                      vreinterpretq_s16_u16(q3u16));
+    d0u8 = vqmovun_s16(q1s16);
+    vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  }
+}
+
+void vp9_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint16x8_t q0u16, q3u16, q10u16;
+  int16x8_t q0s16;
+  uint16x4_t d20u16;
+  uint8x8_t d0u8, d2u8, d30u8;
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d30u8 = vld1_u8(left);
+  d2u8 = vld1_u8(above);
+  q10u16 = vmovl_u8(d30u8);
+  q3u16 = vsubl_u8(d2u8, d0u8);
+  d20u16 = vget_low_u16(q10u16);
+  for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+    q0u16 = vdupq_lane_u16(d20u16, 0);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 1);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 2);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 3);
+    q0s16 = vaddq_s16(vreinterpretq_s16_u16(q3u16),
+                      vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+  }
+}
+
+void vp9_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
+  uint8x16_t q0u8, q1u8;
+  int16x8_t q0s16, q1s16, q8s16, q11s16;
+  uint16x4_t d20u16;
+  uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  for (k = 0; k < 2; k++, left += 8) {
+    d18u8 = vld1_u8(left);
+    q10u16 = vmovl_u8(d18u8);
+    d20u16 = vget_low_u16(q10u16);
+    for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+      q0u16 = vdupq_lane_u16(d20u16, 0);
+      q8u16 = vdupq_lane_u16(d20u16, 1);
+      q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q2u16));
+      q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q3u16));
+      q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                         vreinterpretq_s16_u16(q2u16));
+      q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                        vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d20u16, 2);
+      q8u16 = vdupq_lane_u16(d20u16, 3);
+      q1s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q2u16));
+      q0s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                        vreinterpretq_s16_u16(q3u16));
+      q11s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                         vreinterpretq_s16_u16(q2u16));
+      q8s16 = vaddq_s16(vreinterpretq_s16_u16(q8u16),
+                        vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+    }
+  }
+}
+
+void vp9_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
+  uint8x16_t q0u8, q1u8, q2u8;
+  int16x8_t q12s16, q13s16, q14s16, q15s16;
+  uint16x4_t d6u16;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u8 = vld1q_u8(above + 16);
+  q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
+  q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
+  for (k = 0; k < 4; k++, left += 8) {
+    d26u8 = vld1_u8(left);
+    q3u16 = vmovl_u8(d26u8);
+    d6u16 = vget_low_u16(q3u16);
+    for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
+      q0u16 = vdupq_lane_u16(d6u16, 0);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 1);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 2);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 3);
+      q12s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q8u16));
+      q13s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+    }
+  }
+}
+#endif  // !HAVE_NEON_ASM
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon_asm.asm b/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon_asm.asm
new file mode 100644
index 000000000..14f574a50
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_reconintra_neon_asm.asm
@@ -0,0 +1,630 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+    EXPORT  |vp9_v_predictor_4x4_neon|
+    EXPORT  |vp9_v_predictor_8x8_neon|
+    EXPORT  |vp9_v_predictor_16x16_neon|
+    EXPORT  |vp9_v_predictor_32x32_neon|
+    EXPORT  |vp9_h_predictor_4x4_neon|
+    EXPORT  |vp9_h_predictor_8x8_neon|
+    EXPORT  |vp9_h_predictor_16x16_neon|
+    EXPORT  |vp9_h_predictor_32x32_neon|
+    EXPORT  |vp9_tm_predictor_4x4_neon|
+    EXPORT  |vp9_tm_predictor_8x8_neon|
+    EXPORT  |vp9_tm_predictor_16x16_neon|
+    EXPORT  |vp9_tm_predictor_32x32_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_v_predictor_4x4_neon| PROC
+    vld1.32             {d0[0]}, [r2]
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_v_predictor_4x4_neon|
+
+;void vp9_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_v_predictor_8x8_neon| PROC
+    vld1.8              {d0}, [r2]
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_v_predictor_8x8_neon|
+
+;void vp9_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_v_predictor_16x16_neon| PROC
+    vld1.8              {q0}, [r2]
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_v_predictor_16x16_neon|
+
+;void vp9_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_v_predictor_32x32_neon| PROC
+    vld1.8              {q0, q1}, [r2]
+    mov                 r2, #2
+loop_v
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_v
+    bx                  lr
+    ENDP                ; |vp9_v_predictor_32x32_neon|
+
+;void vp9_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_h_predictor_4x4_neon| PROC
+    vld1.32             {d1[0]}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_h_predictor_4x4_neon|
+
+;void vp9_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_h_predictor_8x8_neon| PROC
+    vld1.64             {d1}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[4]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[5]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[6]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[7]
+    vst1.64             {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_h_predictor_8x8_neon|
+
+;void vp9_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_h_predictor_16x16_neon| PROC
+    vld1.8              {q1}, [r3]
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_h_predictor_16x16_neon|
+
+;void vp9_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_h_predictor_32x32_neon| PROC
+    sub                 r1, r1, #16
+    mov                 r2, #2
+loop_h
+    vld1.8              {q1}, [r3]!
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_h
+    bx                  lr
+    ENDP                ; |vp9_h_predictor_32x32_neon|
+
+;void vp9_tm_predictor_4x4_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_tm_predictor_4x4_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.u8             {d0[]}, [r12]
+
+    ; Load above 4 pixels
+    vld1.32             {d2[0]}, [r2]
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]!
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+
+    ; 3rd row and 4th row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |vp9_tm_predictor_4x4_neon|
+
+;void vp9_tm_predictor_8x8_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_tm_predictor_8x8_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; preload 8 left
+    vld1.8              {d30}, [r3]
+
+    ; Load above 8 pixels
+    vld1.64             {d2}, [r2]
+
+    vmovl.u8            q10, d30
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vdup.16             q0, d20[0]
+    vdup.16             q1, d20[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 3rd row and 4th row
+    vdup.16             q8, d20[2]
+    vdup.16             q9, d20[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    ; 5th row and 6th row
+    vdup.16             q0, d21[0]
+    vdup.16             q1, d21[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 7th row and 8th row
+    vdup.16             q8, d21[2]
+    vdup.16             q9, d21[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    bx                  lr
+    ENDP                ; |vp9_tm_predictor_8x8_neon|
+
+;void vp9_tm_predictor_16x16_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_tm_predictor_16x16_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 8 pixels
+    vld1.8              {q1}, [r2]
+
+    ; preload 8 left into r12
+    vld1.8              {d18}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q2, d2, d0
+    vsubl.u8            q3, d3, d0
+
+    vmovl.u8            q10, d18
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 2 times to process 16 rows.
+    mov                 r2, #2
+
+loop_16x16_neon
+    ; Process two rows.
+    vdup.16             q0, d20[0]
+    vdup.16             q8, d20[1]
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d20[2]                  ; proload next 2 rows data
+    vdup.16             q8, d20[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[0]                  ; proload next 2 rows data
+    vdup.16             q8, d21[1]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[2]                  ; proload next 2 rows data
+    vdup.16             q8, d21[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vld1.8              {d18}, [r3]!                  ; preload 8 left into r12
+    vmovl.u8            q10, d18
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_16x16_neon
+
+    bx                  lr
+    ENDP                ; |vp9_tm_predictor_16x16_neon|
+
+;void vp9_tm_predictor_32x32_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                  const uint8_t *above,
+;                                  const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|vp9_tm_predictor_32x32_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 32 pixels
+    vld1.8              {q1}, [r2]!
+    vld1.8              {q2}, [r2]
+
+    ; preload 8 left pixels
+    vld1.8              {d26}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q8, d2, d0
+    vsubl.u8            q9, d3, d0
+    vsubl.u8            q10, d4, d0
+    vsubl.u8            q11, d5, d0
+
+    vmovl.u8            q3, d26
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 4 times to process 32 rows.
+    mov                 r2, #4
+
+loop_32x32_neon
+    ; Process two rows.
+    vdup.16             q0, d6[0]
+    vdup.16             q2, d6[1]
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q1, d6[2]
+    vdup.16             q2, d6[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q1, q8
+    vadd.s16            q13, q1, q9
+    vadd.s16            q14, q1, q10
+    vadd.s16            q15, q1, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[0]
+    vdup.16             q2, d7[1]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[2]
+    vdup.16             q2, d7[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vld1.8              {d0}, [r3]!                   ; preload 8 left pixels
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vmovl.u8            q3, d0
+    vst1.64             {d24-d27}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_32x32_neon
+
+    bx                  lr
+    ENDP                ; |vp9_tm_predictor_32x32_neon|
+
+    END
diff --git a/media/libvpx/vp9/common/arm/neon/vp9_save_reg_neon.asm b/media/libvpx/vp9/common/arm/neon/vp9_save_reg_neon.asm
new file mode 100644
index 000000000..71c3e7077
--- /dev/null
+++ b/media/libvpx/vp9/common/arm/neon/vp9_save_reg_neon.asm
@@ -0,0 +1,36 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+    EXPORT  |vp9_push_neon|
+    EXPORT  |vp9_pop_neon|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|vp9_push_neon| PROC
+    vst1.i64            {d8, d9, d10, d11}, [r0]!
+    vst1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+|vp9_pop_neon| PROC
+    vld1.i64            {d8, d9, d10, d11}, [r0]!
+    vld1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+    END
+
diff --git a/media/libvpx/vp9/common/vp9_alloccommon.c b/media/libvpx/vp9/common/vp9_alloccommon.c
new file mode 100644
index 000000000..8eda491de
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_alloccommon.c
@@ -0,0 +1,183 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_lock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+void unlock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_unlock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+void vp9_set_mb_mi(VP9_COMMON *cm, int width, int height) {
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2);
+  const int aligned_height = ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2);
+
+  cm->mi_cols = aligned_width >> MI_SIZE_LOG2;
+  cm->mi_rows = aligned_height >> MI_SIZE_LOG2;
+  cm->mi_stride = calc_mi_size(cm->mi_cols);
+
+  cm->mb_cols = (cm->mi_cols + 1) >> 1;
+  cm->mb_rows = (cm->mi_rows + 1) >> 1;
+  cm->MBs = cm->mb_rows * cm->mb_cols;
+}
+
+static int alloc_seg_map(VP9_COMMON *cm, int seg_map_size) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    cm->seg_map_array[i] = (uint8_t *)vpx_calloc(seg_map_size, 1);
+    if (cm->seg_map_array[i] == NULL)
+      return 1;
+  }
+  cm->seg_map_alloc_size = seg_map_size;
+
+  // Init the index.
+  cm->seg_map_idx = 0;
+  cm->prev_seg_map_idx = 1;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  if (!cm->frame_parallel_decode)
+    cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+
+  return 0;
+}
+
+static void free_seg_map(VP9_COMMON *cm) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    vpx_free(cm->seg_map_array[i]);
+    cm->seg_map_array[i] = NULL;
+  }
+
+  cm->current_frame_seg_map = NULL;
+
+  if (!cm->frame_parallel_decode) {
+    cm->last_frame_seg_map = NULL;
+  }
+}
+
+void vp9_free_ref_frame_buffers(BufferPool *pool) {
+  int i;
+
+  for (i = 0; i < FRAME_BUFFERS; ++i) {
+    if (pool->frame_bufs[i].ref_count > 0 &&
+        pool->frame_bufs[i].raw_frame_buffer.data != NULL) {
+      pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer);
+      pool->frame_bufs[i].ref_count = 0;
+    }
+    vpx_free(pool->frame_bufs[i].mvs);
+    pool->frame_bufs[i].mvs = NULL;
+    vp9_free_frame_buffer(&pool->frame_bufs[i].buf);
+  }
+}
+
+void vp9_free_postproc_buffers(VP9_COMMON *cm) {
+#if CONFIG_VP9_POSTPROC
+  vp9_free_frame_buffer(&cm->post_proc_buffer);
+  vp9_free_frame_buffer(&cm->post_proc_buffer_int);
+#else
+  (void)cm;
+#endif
+}
+
+void vp9_free_context_buffers(VP9_COMMON *cm) {
+  cm->free_mi(cm);
+  free_seg_map(cm);
+  vpx_free(cm->above_context);
+  cm->above_context = NULL;
+  vpx_free(cm->above_seg_context);
+  cm->above_seg_context = NULL;
+}
+
+int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) {
+  int new_mi_size;
+
+  vp9_set_mb_mi(cm, width, height);
+  new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
+  if (cm->mi_alloc_size < new_mi_size) {
+    cm->free_mi(cm);
+    if (cm->alloc_mi(cm, new_mi_size))
+      goto fail;
+  }
+
+  if (cm->seg_map_alloc_size < cm->mi_rows * cm->mi_cols) {
+    // Create the segmentation map structure and set to 0.
+    free_seg_map(cm);
+    if (alloc_seg_map(cm, cm->mi_rows * cm->mi_cols))
+      goto fail;
+  }
+
+  if (cm->above_context_alloc_cols < cm->mi_cols) {
+    vpx_free(cm->above_context);
+    cm->above_context = (ENTROPY_CONTEXT *)vpx_calloc(
+        2 * mi_cols_aligned_to_sb(cm->mi_cols) * MAX_MB_PLANE,
+        sizeof(*cm->above_context));
+    if (!cm->above_context) goto fail;
+
+    vpx_free(cm->above_seg_context);
+    cm->above_seg_context = (PARTITION_CONTEXT *)vpx_calloc(
+        mi_cols_aligned_to_sb(cm->mi_cols), sizeof(*cm->above_seg_context));
+    if (!cm->above_seg_context) goto fail;
+    cm->above_context_alloc_cols = cm->mi_cols;
+  }
+
+  return 0;
+
+ fail:
+  vp9_free_context_buffers(cm);
+  return 1;
+}
+
+void vp9_remove_common(VP9_COMMON *cm) {
+  vp9_free_context_buffers(cm);
+
+  vpx_free(cm->fc);
+  cm->fc = NULL;
+  vpx_free(cm->frame_contexts);
+  cm->frame_contexts = NULL;
+}
+
+void vp9_init_context_buffers(VP9_COMMON *cm) {
+  cm->setup_mi(cm);
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
+}
+
+void vp9_swap_current_and_last_seg_map(VP9_COMMON *cm) {
+  // Swap indices.
+  const int tmp = cm->seg_map_idx;
+  cm->seg_map_idx = cm->prev_seg_map_idx;
+  cm->prev_seg_map_idx = tmp;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+}
diff --git a/media/libvpx/vp9/common/vp9_alloccommon.h b/media/libvpx/vp9/common/vp9_alloccommon.h
new file mode 100644
index 000000000..c0e51a6ce
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_alloccommon.h
@@ -0,0 +1,44 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_ALLOCCOMMON_H_
+#define VP9_COMMON_VP9_ALLOCCOMMON_H_
+
+#define INVALID_IDX -1  // Invalid buffer index.
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+struct BufferPool;
+
+void vp9_remove_common(struct VP9Common *cm);
+
+int vp9_alloc_context_buffers(struct VP9Common *cm, int width, int height);
+void vp9_init_context_buffers(struct VP9Common *cm);
+void vp9_free_context_buffers(struct VP9Common *cm);
+
+void vp9_free_ref_frame_buffers(struct BufferPool *pool);
+void vp9_free_postproc_buffers(struct VP9Common *cm);
+
+int vp9_alloc_state_buffers(struct VP9Common *cm, int width, int height);
+void vp9_free_state_buffers(struct VP9Common *cm);
+
+void vp9_set_mb_mi(struct VP9Common *cm, int width, int height);
+
+void vp9_swap_current_and_last_seg_map(struct VP9Common *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ALLOCCOMMON_H_
diff --git a/media/libvpx/vp9/common/vp9_blockd.c b/media/libvpx/vp9/common/vp9_blockd.c
new file mode 100644
index 000000000..b2bb18188
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_blockd.c
@@ -0,0 +1,135 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_blockd.h"
+
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b) {
+  if (b == 0 || b == 2) {
+    if (!left_mi || is_inter_block(&left_mi->mbmi))
+      return DC_PRED;
+
+    return get_y_mode(left_mi, b + 1);
+  } else {
+    assert(b == 1 || b == 3);
+    return cur_mi->bmi[b - 1].as_mode;
+  }
+}
+
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b) {
+  if (b == 0 || b == 1) {
+    if (!above_mi || is_inter_block(&above_mi->mbmi))
+      return DC_PRED;
+
+    return get_y_mode(above_mi, b + 2);
+  } else {
+    assert(b == 2 || b == 3);
+    return cur_mi->bmi[b - 2].as_mode;
+  }
+}
+
+void vp9_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MB_MODE_INFO* mbmi = &xd->mi[0]->mbmi;
+  // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
+  // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
+  // transform size varies per plane, look it up in a common way.
+  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd)
+                                : mbmi->tx_size;
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  const int step = 1 << (tx_size << 1);
+  int i = 0, r, c;
+
+  // If mb_to_right_edge is < 0 we are in a situation in which
+  // the current block size extends into the UMV and we won't
+  // visit the sub blocks that are wholly within the UMV.
+  const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
+      xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+  const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
+      xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += (1 << tx_size)) {
+    for (c = 0; c < num_4x4_w; c += (1 << tx_size)) {
+      // Skip visiting the sub blocks that are wholly within the UMV.
+      if (c < max_blocks_wide)
+        visit(plane, i, plane_bsize, tx_size, arg);
+      i += step;
+    }
+  }
+}
+
+void vp9_foreach_transformed_block(const MACROBLOCKD* const xd,
+                                   BLOCK_SIZE bsize,
+                                   foreach_transformed_block_visitor visit,
+                                   void *arg) {
+  int plane;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+    vp9_foreach_transformed_block_in_plane(xd, bsize, plane, visit, arg);
+}
+
+void vp9_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int has_eob,
+                      int aoff, int loff) {
+  ENTROPY_CONTEXT *const a = pd->above_context + aoff;
+  ENTROPY_CONTEXT *const l = pd->left_context + loff;
+  const int tx_size_in_blocks = 1 << tx_size;
+
+  // above
+  if (has_eob && xd->mb_to_right_edge < 0) {
+    int i;
+    const int blocks_wide = num_4x4_blocks_wide_lookup[plane_bsize] +
+                            (xd->mb_to_right_edge >> (5 + pd->subsampling_x));
+    int above_contexts = tx_size_in_blocks;
+    if (above_contexts + aoff > blocks_wide)
+      above_contexts = blocks_wide - aoff;
+
+    for (i = 0; i < above_contexts; ++i)
+      a[i] = has_eob;
+    for (i = above_contexts; i < tx_size_in_blocks; ++i)
+      a[i] = 0;
+  } else {
+    memset(a, has_eob, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
+  }
+
+  // left
+  if (has_eob && xd->mb_to_bottom_edge < 0) {
+    int i;
+    const int blocks_high = num_4x4_blocks_high_lookup[plane_bsize] +
+                            (xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
+    int left_contexts = tx_size_in_blocks;
+    if (left_contexts + loff > blocks_high)
+      left_contexts = blocks_high - loff;
+
+    for (i = 0; i < left_contexts; ++i)
+      l[i] = has_eob;
+    for (i = left_contexts; i < tx_size_in_blocks; ++i)
+      l[i] = 0;
+  } else {
+    memset(l, has_eob, sizeof(ENTROPY_CONTEXT) * tx_size_in_blocks);
+  }
+}
+
+void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y) {
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].plane_type = i ? PLANE_TYPE_UV : PLANE_TYPE_Y;
+    xd->plane[i].subsampling_x = i ? ss_x : 0;
+    xd->plane[i].subsampling_y = i ? ss_y : 0;
+  }
+}
diff --git a/media/libvpx/vp9/common/vp9_blockd.h b/media/libvpx/vp9/common/vp9_blockd.h
new file mode 100644
index 000000000..e53e15da9
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_blockd.h
@@ -0,0 +1,298 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_BLOCKD_H_
+#define VP9_COMMON_VP9_BLOCKD_H_
+
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx_scale/yv12config.h"
+
+#include "vp9/common/vp9_common_data.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_mv.h"
+#include "vp9/common/vp9_scale.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_MB_PLANE 3
+
+typedef enum {
+  KEY_FRAME = 0,
+  INTER_FRAME = 1,
+  FRAME_TYPES,
+} FRAME_TYPE;
+
+static INLINE int is_inter_mode(PREDICTION_MODE mode) {
+  return mode >= NEARESTMV && mode <= NEWMV;
+}
+
+/* For keyframes, intra block modes are predicted by the (already decoded)
+   modes for the Y blocks to the left and above us; for interframes, there
+   is a single probability table. */
+
+typedef struct {
+  PREDICTION_MODE as_mode;
+  int_mv as_mv[2];  // first, second inter predictor motion vectors
+} b_mode_info;
+
+// Note that the rate-distortion optimization loop, bit-stream writer, and
+// decoder implementation modules critically rely on the enum entry values
+// specified herein. They should be refactored concurrently.
+typedef enum {
+  NONE = -1,
+  INTRA_FRAME = 0,
+  LAST_FRAME = 1,
+  GOLDEN_FRAME = 2,
+  ALTREF_FRAME = 3,
+  MAX_REF_FRAMES = 4
+} MV_REFERENCE_FRAME;
+
+// This structure now relates to 8x8 block regions.
+typedef struct {
+  // Common for both INTER and INTRA blocks
+  BLOCK_SIZE sb_type;
+  PREDICTION_MODE mode;
+  TX_SIZE tx_size;
+  int8_t skip;
+  int8_t segment_id;
+  int8_t seg_id_predicted;  // valid only when temporal_update is enabled
+
+  // Only for INTRA blocks
+  PREDICTION_MODE uv_mode;
+
+  // Only for INTER blocks
+  MV_REFERENCE_FRAME ref_frame[2];
+  int_mv mv[2];
+  int_mv ref_mvs[MAX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  uint8_t mode_context[MAX_REF_FRAMES];
+  INTERP_FILTER interp_filter;
+
+} MB_MODE_INFO;
+
+typedef struct MODE_INFO {
+  MB_MODE_INFO mbmi;
+  b_mode_info bmi[4];
+} MODE_INFO;
+
+static INLINE PREDICTION_MODE get_y_mode(const MODE_INFO *mi, int block) {
+  return mi->mbmi.sb_type < BLOCK_8X8 ? mi->bmi[block].as_mode
+                                      : mi->mbmi.mode;
+}
+
+static INLINE int is_inter_block(const MB_MODE_INFO *mbmi) {
+  return mbmi->ref_frame[0] > INTRA_FRAME;
+}
+
+static INLINE int has_second_ref(const MB_MODE_INFO *mbmi) {
+  return mbmi->ref_frame[1] > INTRA_FRAME;
+}
+
+PREDICTION_MODE vp9_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b);
+
+PREDICTION_MODE vp9_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b);
+
+enum mv_precision {
+  MV_PRECISION_Q3,
+  MV_PRECISION_Q4
+};
+
+struct buf_2d {
+  uint8_t *buf;
+  int stride;
+};
+
+struct macroblockd_plane {
+  tran_low_t *dqcoeff;
+  PLANE_TYPE plane_type;
+  int subsampling_x;
+  int subsampling_y;
+  struct buf_2d dst;
+  struct buf_2d pre[2];
+  ENTROPY_CONTEXT *above_context;
+  ENTROPY_CONTEXT *left_context;
+  int16_t seg_dequant[MAX_SEGMENTS][2];
+
+  // encoder
+  const int16_t *dequant;
+};
+
+#define BLOCK_OFFSET(x, i) ((x) + (i) * 16)
+
+typedef struct RefBuffer {
+  // TODO(dkovalev): idx is not really required and should be removed, now it
+  // is used in vp9_onyxd_if.c
+  int idx;
+  YV12_BUFFER_CONFIG *buf;
+  struct scale_factors sf;
+} RefBuffer;
+
+typedef struct macroblockd {
+  struct macroblockd_plane plane[MAX_MB_PLANE];
+  FRAME_COUNTS *counts;
+  int mi_stride;
+
+  MODE_INFO **mi;
+  MODE_INFO *left_mi;
+  MODE_INFO *above_mi;
+  MB_MODE_INFO *left_mbmi;
+  MB_MODE_INFO *above_mbmi;
+
+  int up_available;
+  int left_available;
+
+  /* Distance of MB away from frame edges */
+  int mb_to_left_edge;
+  int mb_to_right_edge;
+  int mb_to_top_edge;
+  int mb_to_bottom_edge;
+
+  FRAME_CONTEXT *fc;
+  int frame_parallel_decoding_mode;
+
+  /* pointers to reference frames */
+  RefBuffer *block_refs[2];
+
+  /* pointer to current frame */
+  const YV12_BUFFER_CONFIG *cur_buf;
+
+  ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
+  ENTROPY_CONTEXT left_context[MAX_MB_PLANE][16];
+
+  PARTITION_CONTEXT *above_seg_context;
+  PARTITION_CONTEXT left_seg_context[8];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  /* Bit depth: 8, 10, 12 */
+  int bd;
+#endif
+
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[64 * 64]);
+
+  int lossless;
+  int corrupted;
+
+  struct vpx_internal_error_info *error_info;
+} MACROBLOCKD;
+
+static INLINE BLOCK_SIZE get_subsize(BLOCK_SIZE bsize,
+                                     PARTITION_TYPE partition) {
+  return subsize_lookup[partition][bsize];
+}
+
+extern const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES];
+
+static INLINE TX_TYPE get_tx_type(PLANE_TYPE plane_type,
+                                  const MACROBLOCKD *xd) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  if (plane_type != PLANE_TYPE_Y || xd->lossless || is_inter_block(mbmi))
+    return DCT_DCT;
+
+  return intra_mode_to_tx_type_lookup[mbmi->mode];
+}
+
+static INLINE TX_TYPE get_tx_type_4x4(PLANE_TYPE plane_type,
+                                      const MACROBLOCKD *xd, int ib) {
+  const MODE_INFO *const mi = xd->mi[0];
+
+  if (plane_type != PLANE_TYPE_Y || xd->lossless || is_inter_block(&mi->mbmi))
+    return DCT_DCT;
+
+  return intra_mode_to_tx_type_lookup[get_y_mode(mi, ib)];
+}
+
+void vp9_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
+
+static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize,
+                                          int xss, int yss) {
+  if (bsize < BLOCK_8X8) {
+    return TX_4X4;
+  } else {
+    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][xss][yss];
+    return MIN(y_tx_size, max_txsize_lookup[plane_bsize]);
+  }
+}
+
+static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi,
+                                     const struct macroblockd_plane *pd) {
+  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type, pd->subsampling_x,
+                             pd->subsampling_y);
+}
+
+static INLINE BLOCK_SIZE get_plane_block_size(BLOCK_SIZE bsize,
+    const struct macroblockd_plane *pd) {
+  return ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
+}
+
+static INLINE void reset_skip_context(MACROBLOCKD *xd, BLOCK_SIZE bsize) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    memset(pd->above_context, 0,
+           sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide_lookup[plane_bsize]);
+    memset(pd->left_context, 0,
+           sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high_lookup[plane_bsize]);
+  }
+}
+
+static INLINE const vp9_prob *get_y_mode_probs(const MODE_INFO *mi,
+                                               const MODE_INFO *above_mi,
+                                               const MODE_INFO *left_mi,
+                                               int block) {
+  const PREDICTION_MODE above = vp9_above_block_mode(mi, above_mi, block);
+  const PREDICTION_MODE left = vp9_left_block_mode(mi, left_mi, block);
+  return vp9_kf_y_mode_prob[above][left];
+}
+
+typedef void (*foreach_transformed_block_visitor)(int plane, int block,
+                                                  BLOCK_SIZE plane_bsize,
+                                                  TX_SIZE tx_size,
+                                                  void *arg);
+
+void vp9_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg);
+
+
+void vp9_foreach_transformed_block(
+    const MACROBLOCKD* const xd, BLOCK_SIZE bsize,
+    foreach_transformed_block_visitor visit, void *arg);
+
+static INLINE void txfrm_block_to_raster_xy(BLOCK_SIZE plane_bsize,
+                                            TX_SIZE tx_size, int block,
+                                            int *x, int *y) {
+  const int bwl = b_width_log2_lookup[plane_bsize];
+  const int tx_cols_log2 = bwl - tx_size;
+  const int tx_cols = 1 << tx_cols_log2;
+  const int raster_mb = block >> (tx_size << 1);
+  *x = (raster_mb & (tx_cols - 1)) << tx_size;
+  *y = (raster_mb >> tx_cols_log2) << tx_size;
+}
+
+void vp9_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int has_eob,
+                      int aoff, int loff);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_BLOCKD_H_
diff --git a/media/libvpx/vp9/common/vp9_common.h b/media/libvpx/vp9/common/vp9_common.h
new file mode 100644
index 000000000..9c2d7791e
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_common.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_COMMON_H_
+#define VP9_COMMON_VP9_COMMON_H_
+
+/* Interface header for common constant data structures and lookup tables */
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MIN(x, y) (((x) < (y)) ? (x) : (y))
+#define MAX(x, y) (((x) > (y)) ? (x) : (y))
+
+// Only need this for fixed-size arrays, for structs just assign.
+#define vp9_copy(dest, src) {            \
+    assert(sizeof(dest) == sizeof(src)); \
+    memcpy(dest, src, sizeof(src));  \
+  }
+
+// Use this for variably-sized arrays.
+#define vp9_copy_array(dest, src, n) {       \
+    assert(sizeof(*dest) == sizeof(*src));   \
+    memcpy(dest, src, n * sizeof(*src)); \
+  }
+
+#define vp9_zero(dest) memset(&(dest), 0, sizeof(dest))
+#define vp9_zero_array(dest, n) memset(dest, 0, n * sizeof(*dest))
+
+static INLINE uint8_t clip_pixel(int val) {
+  return (val > 255) ? 255 : (val < 0) ? 0 : val;
+}
+
+static INLINE int clamp(int value, int low, int high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+static INLINE double fclamp(double value, double low, double high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+static INLINE int get_unsigned_bits(unsigned int num_values) {
+  return num_values > 0 ? get_msb(num_values) + 1 : 0;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE uint16_t clip_pixel_highbd(int val, int bd) {
+  switch (bd) {
+    case 8:
+    default:
+      return (uint16_t)clamp(val, 0, 255);
+    case 10:
+      return (uint16_t)clamp(val, 0, 1023);
+    case 12:
+      return (uint16_t)clamp(val, 0, 4095);
+  }
+}
+
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+
+#else
+
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_DEBUG
+#define CHECK_MEM_ERROR(cm, lval, expr) do { \
+  lval = (expr); \
+  if (!lval) \
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, \
+                       "Failed to allocate "#lval" at %s:%d", \
+                       __FILE__, __LINE__); \
+  } while (0)
+#else
+#define CHECK_MEM_ERROR(cm, lval, expr) do { \
+  lval = (expr); \
+  if (!lval) \
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, \
+                       "Failed to allocate "#lval); \
+  } while (0)
+#endif
+
+#define VP9_SYNC_CODE_0 0x49
+#define VP9_SYNC_CODE_1 0x83
+#define VP9_SYNC_CODE_2 0x42
+
+#define VP9_FRAME_MARKER 0x2
+
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_COMMON_H_
diff --git a/media/libvpx/vp9/common/vp9_common_data.c b/media/libvpx/vp9/common/vp9_common_data.c
new file mode 100644
index 000000000..2aaa009fa
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_common_data.c
@@ -0,0 +1,160 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common_data.h"
+
+// Log 2 conversion lookup tables for block width and height
+const int b_width_log2_lookup[BLOCK_SIZES] =
+  {0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4};
+const int b_height_log2_lookup[BLOCK_SIZES] =
+  {0, 1, 0, 1, 2, 1, 2, 3, 2, 3, 4, 3, 4};
+const int num_4x4_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 8, 16, 16};
+const int num_4x4_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 16, 8, 16};
+// Log 2 conversion lookup tables for modeinfo width and height
+const int mi_width_log2_lookup[BLOCK_SIZES] =
+  {0, 0, 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3};
+const int num_8x8_blocks_wide_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8};
+const int num_8x8_blocks_high_lookup[BLOCK_SIZES] =
+  {1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8};
+
+// MIN(3, MIN(b_width_log2(bsize), b_height_log2(bsize)))
+const int size_group_lookup[BLOCK_SIZES] =
+  {0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3};
+
+const int num_pels_log2_lookup[BLOCK_SIZES] =
+  {4, 5, 5, 6, 7, 7, 8, 9, 9, 10, 11, 11, 12};
+
+const PARTITION_TYPE partition_lookup[][BLOCK_SIZES] = {
+  {  // 4X4
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_NONE, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID
+  }, {  // 8X8
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_VERT, PARTITION_HORZ, PARTITION_NONE,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 16X16
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_VERT, PARTITION_HORZ, PARTITION_NONE, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 32X32
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_VERT,
+    PARTITION_HORZ, PARTITION_NONE, PARTITION_INVALID,
+    PARTITION_INVALID, PARTITION_INVALID
+  }, {  // 64X64
+    // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_SPLIT,
+    PARTITION_SPLIT, PARTITION_SPLIT, PARTITION_VERT, PARTITION_HORZ,
+    PARTITION_NONE
+  }
+};
+
+const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES] = {
+  {     // PARTITION_NONE
+    BLOCK_4X4,   BLOCK_4X8,   BLOCK_8X4,
+    BLOCK_8X8,   BLOCK_8X16,  BLOCK_16X8,
+    BLOCK_16X16, BLOCK_16X32, BLOCK_32X16,
+    BLOCK_32X32, BLOCK_32X64, BLOCK_64X32,
+    BLOCK_64X64,
+  }, {  // PARTITION_HORZ
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X4,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X8,    BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X16,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_64X32,
+  }, {  // PARTITION_VERT
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_4X8,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X16,    BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X32,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X64,
+  }, {  // PARTITION_SPLIT
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_4X4,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_8X8,     BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_16X16,   BLOCK_INVALID, BLOCK_INVALID,
+    BLOCK_32X32,
+  }
+};
+
+const TX_SIZE max_txsize_lookup[BLOCK_SIZES] = {
+  TX_4X4,   TX_4X4,   TX_4X4,
+  TX_8X8,   TX_8X8,   TX_8X8,
+  TX_16X16, TX_16X16, TX_16X16,
+  TX_32X32, TX_32X32, TX_32X32, TX_32X32
+};
+
+const BLOCK_SIZE txsize_to_bsize[TX_SIZES] = {
+    BLOCK_4X4,  // TX_4X4
+    BLOCK_8X8,  // TX_8X8
+    BLOCK_16X16,  // TX_16X16
+    BLOCK_32X32,  // TX_32X32
+};
+
+const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES] = {
+  TX_4X4,  // ONLY_4X4
+  TX_8X8,  // ALLOW_8X8
+  TX_16X16,  // ALLOW_16X16
+  TX_32X32,  // ALLOW_32X32
+  TX_32X32,  // TX_MODE_SELECT
+};
+
+const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2] = {
+//  ss_x == 0    ss_x == 0        ss_x == 1      ss_x == 1
+//  ss_y == 0    ss_y == 1        ss_y == 0      ss_y == 1
+  {{BLOCK_4X4,   BLOCK_INVALID}, {BLOCK_INVALID, BLOCK_INVALID}},
+  {{BLOCK_4X8,   BLOCK_4X4},     {BLOCK_INVALID, BLOCK_INVALID}},
+  {{BLOCK_8X4,   BLOCK_INVALID}, {BLOCK_4X4,     BLOCK_INVALID}},
+  {{BLOCK_8X8,   BLOCK_8X4},     {BLOCK_4X8,     BLOCK_4X4}},
+  {{BLOCK_8X16,  BLOCK_8X8},     {BLOCK_INVALID, BLOCK_4X8}},
+  {{BLOCK_16X8,  BLOCK_INVALID}, {BLOCK_8X8,     BLOCK_8X4}},
+  {{BLOCK_16X16, BLOCK_16X8},    {BLOCK_8X16,    BLOCK_8X8}},
+  {{BLOCK_16X32, BLOCK_16X16},   {BLOCK_INVALID, BLOCK_8X16}},
+  {{BLOCK_32X16, BLOCK_INVALID}, {BLOCK_16X16,   BLOCK_16X8}},
+  {{BLOCK_32X32, BLOCK_32X16},   {BLOCK_16X32,   BLOCK_16X16}},
+  {{BLOCK_32X64, BLOCK_32X32},   {BLOCK_INVALID, BLOCK_16X32}},
+  {{BLOCK_64X32, BLOCK_INVALID}, {BLOCK_32X32,   BLOCK_32X16}},
+  {{BLOCK_64X64, BLOCK_64X32},   {BLOCK_32X64,   BLOCK_32X32}},
+};
+
+// Generates 4 bit field in which each bit set to 1 represents
+// a blocksize partition  1111 means we split 64x64, 32x32, 16x16
+// and 8x8.  1000 means we just split the 64x64 to 32x32
+const struct {
+  PARTITION_CONTEXT above;
+  PARTITION_CONTEXT left;
+} partition_context_lookup[BLOCK_SIZES]= {
+  {15, 15},  // 4X4   - {0b1111, 0b1111}
+  {15, 14},  // 4X8   - {0b1111, 0b1110}
+  {14, 15},  // 8X4   - {0b1110, 0b1111}
+  {14, 14},  // 8X8   - {0b1110, 0b1110}
+  {14, 12},  // 8X16  - {0b1110, 0b1100}
+  {12, 14},  // 16X8  - {0b1100, 0b1110}
+  {12, 12},  // 16X16 - {0b1100, 0b1100}
+  {12, 8 },  // 16X32 - {0b1100, 0b1000}
+  {8,  12},  // 32X16 - {0b1000, 0b1100}
+  {8,  8 },  // 32X32 - {0b1000, 0b1000}
+  {8,  0 },  // 32X64 - {0b1000, 0b0000}
+  {0,  8 },  // 64X32 - {0b0000, 0b1000}
+  {0,  0 },  // 64X64 - {0b0000, 0b0000}
+};
diff --git a/media/libvpx/vp9/common/vp9_common_data.h b/media/libvpx/vp9/common/vp9_common_data.h
new file mode 100644
index 000000000..a06c9bed8
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_common_data.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_COMMON_DATA_H_
+#define VP9_COMMON_VP9_COMMON_DATA_H_
+
+#include "vp9/common/vp9_enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const int b_width_log2_lookup[BLOCK_SIZES];
+extern const int b_height_log2_lookup[BLOCK_SIZES];
+extern const int mi_width_log2_lookup[BLOCK_SIZES];
+extern const int num_8x8_blocks_wide_lookup[BLOCK_SIZES];
+extern const int num_8x8_blocks_high_lookup[BLOCK_SIZES];
+extern const int num_4x4_blocks_high_lookup[BLOCK_SIZES];
+extern const int num_4x4_blocks_wide_lookup[BLOCK_SIZES];
+extern const int size_group_lookup[BLOCK_SIZES];
+extern const int num_pels_log2_lookup[BLOCK_SIZES];
+extern const PARTITION_TYPE partition_lookup[][BLOCK_SIZES];
+extern const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES];
+extern const TX_SIZE max_txsize_lookup[BLOCK_SIZES];
+extern const BLOCK_SIZE txsize_to_bsize[TX_SIZES];
+extern const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES];
+extern const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_COMMON_DATA_H_
diff --git a/media/libvpx/vp9/common/vp9_convolve.c b/media/libvpx/vp9/common/vp9_convolve.c
new file mode 100644
index 000000000..90e337fd6
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_convolve.c
@@ -0,0 +1,557 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_convolve.h"
+#include "vp9/common/vp9_filter.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const InterpKernel *x_filters,
+                           int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const InterpKernel *x_filters,
+                               int x0_q4, int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const InterpKernel *y_filters,
+                          int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const InterpKernel *y_filters,
+                              int y0_q4, int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const InterpKernel *const x_filters,
+                     int x0_q4, int x_step_q4,
+                     const InterpKernel *const y_filters,
+                     int y0_q4, int y_step_q4,
+                     int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[135 * 64];
+  int intermediate_height =
+          (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
+                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
+  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
+                y_filters, y0_q4, y_step_q4, w, h);
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4,
+                           int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                 x0_q4, x_step_q4, w, h);
+}
+
+void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                     x0_q4, x_step_q4, w, h);
+}
+
+void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int x_step_q4,
+                          const int16_t *filter_y, int y_step_q4,
+                          int w, int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                y0_q4, y_step_q4, w, h);
+}
+
+void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                    y0_q4, y_step_q4, w, h);
+}
+
+void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                     uint8_t *dst, ptrdiff_t dst_stride,
+                     const int16_t *filter_x, int x_step_q4,
+                     const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride,
+           filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int x_step_q4,
+                         const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(16, uint8_t, temp[64 * 64]);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vp9_convolve8_c(src, src_stride, temp, 64,
+                  filter_x, x_step_q4, filter_y, y_step_q4, w, h);
+  vp9_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
+}
+
+void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
+                         uint8_t *dst, ptrdiff_t dst_stride,
+                         const int16_t *filter_x, int filter_x_stride,
+                         const int16_t *filter_y, int filter_y_stride,
+                         int w, int h) {
+  int r;
+
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                        uint8_t *dst, ptrdiff_t dst_stride,
+                        const int16_t *filter_x, int filter_x_stride,
+                        const int16_t *filter_y, int filter_y_stride,
+                        int w, int h) {
+  int x, y;
+
+  (void)filter_x;  (void)filter_x_stride;
+  (void)filter_y;  (void)filter_y_stride;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x)
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                  uint8_t *dst8, ptrdiff_t dst_stride,
+                                  const InterpKernel *x_filters,
+                                  int x0_q4, int x_step_q4,
+                                  int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                      uint8_t *dst8, ptrdiff_t dst_stride,
+                                      const InterpKernel *x_filters,
+                                      int x0_q4, int x_step_q4,
+                                      int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                 uint8_t *dst8, ptrdiff_t dst_stride,
+                                 const InterpKernel *y_filters,
+                                 int y0_q4, int y_step_q4, int w, int h,
+                                 int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel_highbd(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                     uint8_t *dst8, ptrdiff_t dst_stride,
+                                     const InterpKernel *y_filters,
+                                     int y0_q4, int y_step_q4, int w, int h,
+                                     int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const InterpKernel *const x_filters,
+                            int x0_q4, int x_step_q4,
+                            const InterpKernel *const y_filters,
+                            int y0_q4, int y_step_q4,
+                            int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[64 * 135];
+  int intermediate_height =
+          (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                        src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                        x_filters, x0_q4, x_step_q4, w,
+                        intermediate_height, bd);
+  highbd_convolve_vert(CONVERT_TO_BYTEPTR(temp) + 64 * (SUBPEL_TAPS / 2 - 1),
+                       64, dst, dst_stride, y_filters, y0_q4, y_step_q4,
+                       w, h, bd);
+}
+
+
+void vp9_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4,
+                                  int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                        x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const int16_t *filter_x, int x_step_q4,
+                                      const int16_t *filter_y, int y_step_q4,
+                                      int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                            x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4,
+                                 int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                       y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                           y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4,
+                            int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve(src, src_stride, dst, dst_stride,
+                  filters_x, x0_q4, x_step_q4,
+                  filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4,
+                                int w, int h, int bd) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED(16, uint16_t, temp[64 * 64]);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vp9_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+  vp9_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), 64, dst, dst_stride,
+                            NULL, 0, NULL, 0, w, h, bd);
+}
+
+void vp9_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+                                uint8_t *dst8, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int filter_x_stride,
+                                const int16_t *filter_y, int filter_y_stride,
+                                int w, int h, int bd) {
+  int r;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w * sizeof(uint16_t));
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vp9_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+                               uint8_t *dst8, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int filter_x_stride,
+                               const int16_t *filter_y, int filter_y_stride,
+                               int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+#endif
diff --git a/media/libvpx/vp9/common/vp9_convolve.h b/media/libvpx/vp9/common/vp9_convolve.h
new file mode 100644
index 000000000..8b044c897
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_convolve.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP9_COMMON_VP9_CONVOLVE_H_
+#define VP9_COMMON_VP9_CONVOLVE_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*highbd_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_CONVOLVE_H_
diff --git a/media/libvpx/vp9/common/vp9_debugmodes.c b/media/libvpx/vp9/common/vp9_debugmodes.c
new file mode 100644
index 000000000..3d80103d2
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_debugmodes.c
@@ -0,0 +1,91 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+static void log_frame_info(VP9_COMMON *cm, const char *str, FILE *f) {
+  fprintf(f, "%s", str);
+  fprintf(f, "(Frame %d, Show:%d, Q:%d): \n", cm->current_video_frame,
+          cm->show_frame, cm->base_qindex);
+}
+/* This function dereferences a pointer to the mbmi structure
+ * and uses the passed in member offset to print out the value of an integer
+ * for each mbmi member value in the mi structure.
+ */
+static void print_mi_data(VP9_COMMON *cm, FILE *file, const char *descriptor,
+                          size_t member_offset) {
+  int mi_row, mi_col;
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+  char prefix = descriptor[0];
+
+  log_frame_info(cm, descriptor, file);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(file, "%c ", prefix);
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(file, "%2d ",
+              *((int*) ((char *) (&mi[0]->mbmi) +
+                                  member_offset)));
+      mi++;
+    }
+    fprintf(file, "\n");
+    mi += 8;
+  }
+  fprintf(file, "\n");
+}
+
+void vp9_print_modes_and_motion_vectors(VP9_COMMON *cm, const char *file) {
+  int mi_row;
+  int mi_col;
+  FILE *mvs = fopen(file, "a");
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+
+  print_mi_data(cm, mvs, "Partitions:", offsetof(MB_MODE_INFO, sb_type));
+  print_mi_data(cm, mvs, "Modes:", offsetof(MB_MODE_INFO, mode));
+  print_mi_data(cm, mvs, "Ref frame:", offsetof(MB_MODE_INFO, ref_frame[0]));
+  print_mi_data(cm, mvs, "Transform:", offsetof(MB_MODE_INFO, tx_size));
+  print_mi_data(cm, mvs, "UV Modes:", offsetof(MB_MODE_INFO, uv_mode));
+
+  // output skip infomation.
+  log_frame_info(cm, "Skips:", mvs);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "S ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%2d ", mi[0]->mbmi.skip);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  // output motion vectors.
+  log_frame_info(cm, "Vectors ", mvs);
+  mi = cm->mi_grid_visible;
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "V ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%4d:%4d ", mi[0]->mbmi.mv[0].as_mv.row,
+                               mi[0]->mbmi.mv[0].as_mv.col);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  fclose(mvs);
+}
diff --git a/media/libvpx/vp9/common/vp9_entropy.c b/media/libvpx/vp9/common/vp9_entropy.c
new file mode 100644
index 000000000..a2584e8da
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropy.c
@@ -0,0 +1,823 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx/vpx_integer.h"
+
+// Unconstrained Node Tree
+const vp9_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  2, 6,                                // 0 = LOW_VAL
+  -TWO_TOKEN, 4,                       // 1 = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 2 = THREE
+  8, 10,                               // 3 = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 4 = CAT_ONE
+  12, 14,                              // 5 = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 6 = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 7 = CAT_FIVE
+};
+
+const vp9_prob vp9_cat1_prob[] = { 159 };
+const vp9_prob vp9_cat2_prob[] = { 165, 145 };
+const vp9_prob vp9_cat3_prob[] = { 173, 148, 140 };
+const vp9_prob vp9_cat4_prob[] = { 176, 155, 140, 135 };
+const vp9_prob vp9_cat5_prob[] = { 180, 157, 141, 134, 130 };
+const vp9_prob vp9_cat6_prob[] = {
+    254, 254, 254, 252, 249, 243, 230, 196, 177, 153, 140, 133, 130, 129
+};
+#if CONFIG_VP9_HIGHBITDEPTH
+const vp9_prob vp9_cat1_prob_high10[] = { 159 };
+const vp9_prob vp9_cat2_prob_high10[] = { 165, 145 };
+const vp9_prob vp9_cat3_prob_high10[] = { 173, 148, 140 };
+const vp9_prob vp9_cat4_prob_high10[] = { 176, 155, 140, 135 };
+const vp9_prob vp9_cat5_prob_high10[] = { 180, 157, 141, 134, 130 };
+const vp9_prob vp9_cat6_prob_high10[] = {
+    255, 255, 254, 254, 254, 252, 249, 243,
+    230, 196, 177, 153, 140, 133, 130, 129
+};
+const vp9_prob vp9_cat1_prob_high12[] = { 159 };
+const vp9_prob vp9_cat2_prob_high12[] = { 165, 145 };
+const vp9_prob vp9_cat3_prob_high12[] = { 173, 148, 140 };
+const vp9_prob vp9_cat4_prob_high12[] = { 176, 155, 140, 135 };
+const vp9_prob vp9_cat5_prob_high12[] = { 180, 157, 141, 134, 130 };
+const vp9_prob vp9_cat6_prob_high12[] = {
+    255, 255, 255, 255, 254, 254, 254, 252, 249,
+    243, 230, 196, 177, 153, 140, 133, 130, 129
+};
+#endif
+
+const uint8_t vp9_coefband_trans_8x8plus[1024] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4,
+  4, 4, 4, 4, 4, 5,
+  // beyond MAXBAND_INDEX+1 all values are filled as 5
+                    5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+};
+
+const uint8_t vp9_coefband_trans_4x4[16] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5,
+};
+
+const uint8_t vp9_pt_energy_class[ENTROPY_TOKENS] = {
+  0, 1, 2, 3, 3, 4, 4, 5, 5, 5, 5, 5
+};
+
+const vp9_tree_index vp9_coefmodel_tree[TREE_SIZE(UNCONSTRAINED_NODES + 1)] = {
+  -EOB_MODEL_TOKEN, 2,
+  -ZERO_TOKEN, 4,
+  -ONE_TOKEN, -TWO_TOKEN,
+};
+
+// Model obtained from a 2-sided zero-centerd distribuition derived
+// from a Pareto distribution. The cdf of the distribution is:
+// cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
+//
+// For a given beta and a given probablity of the 1-node, the alpha
+// is first solved, and then the {alpha, beta} pair is used to generate
+// the probabilities for the rest of the nodes.
+
+// beta = 8
+
+// Every odd line in this table can be generated from the even lines
+// by averaging :
+// vp9_pareto8_full[l][node] = (vp9_pareto8_full[l-1][node] +
+//                              vp9_pareto8_full[l+1][node] ) >> 1;
+const vp9_prob vp9_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES] = {
+  {  3,  86, 128,   6,  86,  23,  88,  29},
+  {  6,  86, 128,  11,  87,  42,  91,  52},
+  {  9,  86, 129,  17,  88,  61,  94,  76},
+  { 12,  86, 129,  22,  88,  77,  97,  93},
+  { 15,  87, 129,  28,  89,  93, 100, 110},
+  { 17,  87, 129,  33,  90, 105, 103, 123},
+  { 20,  88, 130,  38,  91, 118, 106, 136},
+  { 23,  88, 130,  43,  91, 128, 108, 146},
+  { 26,  89, 131,  48,  92, 139, 111, 156},
+  { 28,  89, 131,  53,  93, 147, 114, 163},
+  { 31,  90, 131,  58,  94, 156, 117, 171},
+  { 34,  90, 131,  62,  94, 163, 119, 177},
+  { 37,  90, 132,  66,  95, 171, 122, 184},
+  { 39,  90, 132,  70,  96, 177, 124, 189},
+  { 42,  91, 132,  75,  97, 183, 127, 194},
+  { 44,  91, 132,  79,  97, 188, 129, 198},
+  { 47,  92, 133,  83,  98, 193, 132, 202},
+  { 49,  92, 133,  86,  99, 197, 134, 205},
+  { 52,  93, 133,  90, 100, 201, 137, 208},
+  { 54,  93, 133,  94, 100, 204, 139, 211},
+  { 57,  94, 134,  98, 101, 208, 142, 214},
+  { 59,  94, 134, 101, 102, 211, 144, 216},
+  { 62,  94, 135, 105, 103, 214, 146, 218},
+  { 64,  94, 135, 108, 103, 216, 148, 220},
+  { 66,  95, 135, 111, 104, 219, 151, 222},
+  { 68,  95, 135, 114, 105, 221, 153, 223},
+  { 71,  96, 136, 117, 106, 224, 155, 225},
+  { 73,  96, 136, 120, 106, 225, 157, 226},
+  { 76,  97, 136, 123, 107, 227, 159, 228},
+  { 78,  97, 136, 126, 108, 229, 160, 229},
+  { 80,  98, 137, 129, 109, 231, 162, 231},
+  { 82,  98, 137, 131, 109, 232, 164, 232},
+  { 84,  98, 138, 134, 110, 234, 166, 233},
+  { 86,  98, 138, 137, 111, 235, 168, 234},
+  { 89,  99, 138, 140, 112, 236, 170, 235},
+  { 91,  99, 138, 142, 112, 237, 171, 235},
+  { 93, 100, 139, 145, 113, 238, 173, 236},
+  { 95, 100, 139, 147, 114, 239, 174, 237},
+  { 97, 101, 140, 149, 115, 240, 176, 238},
+  { 99, 101, 140, 151, 115, 241, 177, 238},
+  {101, 102, 140, 154, 116, 242, 179, 239},
+  {103, 102, 140, 156, 117, 242, 180, 239},
+  {105, 103, 141, 158, 118, 243, 182, 240},
+  {107, 103, 141, 160, 118, 243, 183, 240},
+  {109, 104, 141, 162, 119, 244, 185, 241},
+  {111, 104, 141, 164, 119, 244, 186, 241},
+  {113, 104, 142, 166, 120, 245, 187, 242},
+  {114, 104, 142, 168, 121, 245, 188, 242},
+  {116, 105, 143, 170, 122, 246, 190, 243},
+  {118, 105, 143, 171, 122, 246, 191, 243},
+  {120, 106, 143, 173, 123, 247, 192, 244},
+  {121, 106, 143, 175, 124, 247, 193, 244},
+  {123, 107, 144, 177, 125, 248, 195, 244},
+  {125, 107, 144, 178, 125, 248, 196, 244},
+  {127, 108, 145, 180, 126, 249, 197, 245},
+  {128, 108, 145, 181, 127, 249, 198, 245},
+  {130, 109, 145, 183, 128, 249, 199, 245},
+  {132, 109, 145, 184, 128, 249, 200, 245},
+  {134, 110, 146, 186, 129, 250, 201, 246},
+  {135, 110, 146, 187, 130, 250, 202, 246},
+  {137, 111, 147, 189, 131, 251, 203, 246},
+  {138, 111, 147, 190, 131, 251, 204, 246},
+  {140, 112, 147, 192, 132, 251, 205, 247},
+  {141, 112, 147, 193, 132, 251, 206, 247},
+  {143, 113, 148, 194, 133, 251, 207, 247},
+  {144, 113, 148, 195, 134, 251, 207, 247},
+  {146, 114, 149, 197, 135, 252, 208, 248},
+  {147, 114, 149, 198, 135, 252, 209, 248},
+  {149, 115, 149, 199, 136, 252, 210, 248},
+  {150, 115, 149, 200, 137, 252, 210, 248},
+  {152, 115, 150, 201, 138, 252, 211, 248},
+  {153, 115, 150, 202, 138, 252, 212, 248},
+  {155, 116, 151, 204, 139, 253, 213, 249},
+  {156, 116, 151, 205, 139, 253, 213, 249},
+  {158, 117, 151, 206, 140, 253, 214, 249},
+  {159, 117, 151, 207, 141, 253, 215, 249},
+  {161, 118, 152, 208, 142, 253, 216, 249},
+  {162, 118, 152, 209, 142, 253, 216, 249},
+  {163, 119, 153, 210, 143, 253, 217, 249},
+  {164, 119, 153, 211, 143, 253, 217, 249},
+  {166, 120, 153, 212, 144, 254, 218, 250},
+  {167, 120, 153, 212, 145, 254, 219, 250},
+  {168, 121, 154, 213, 146, 254, 220, 250},
+  {169, 121, 154, 214, 146, 254, 220, 250},
+  {171, 122, 155, 215, 147, 254, 221, 250},
+  {172, 122, 155, 216, 147, 254, 221, 250},
+  {173, 123, 155, 217, 148, 254, 222, 250},
+  {174, 123, 155, 217, 149, 254, 222, 250},
+  {176, 124, 156, 218, 150, 254, 223, 250},
+  {177, 124, 156, 219, 150, 254, 223, 250},
+  {178, 125, 157, 220, 151, 254, 224, 251},
+  {179, 125, 157, 220, 151, 254, 224, 251},
+  {180, 126, 157, 221, 152, 254, 225, 251},
+  {181, 126, 157, 221, 152, 254, 225, 251},
+  {183, 127, 158, 222, 153, 254, 226, 251},
+  {184, 127, 158, 223, 154, 254, 226, 251},
+  {185, 128, 159, 224, 155, 255, 227, 251},
+  {186, 128, 159, 224, 155, 255, 227, 251},
+  {187, 129, 160, 225, 156, 255, 228, 251},
+  {188, 130, 160, 225, 156, 255, 228, 251},
+  {189, 131, 160, 226, 157, 255, 228, 251},
+  {190, 131, 160, 226, 158, 255, 228, 251},
+  {191, 132, 161, 227, 159, 255, 229, 251},
+  {192, 132, 161, 227, 159, 255, 229, 251},
+  {193, 133, 162, 228, 160, 255, 230, 252},
+  {194, 133, 162, 229, 160, 255, 230, 252},
+  {195, 134, 163, 230, 161, 255, 231, 252},
+  {196, 134, 163, 230, 161, 255, 231, 252},
+  {197, 135, 163, 231, 162, 255, 231, 252},
+  {198, 135, 163, 231, 162, 255, 231, 252},
+  {199, 136, 164, 232, 163, 255, 232, 252},
+  {200, 136, 164, 232, 164, 255, 232, 252},
+  {201, 137, 165, 233, 165, 255, 233, 252},
+  {201, 137, 165, 233, 165, 255, 233, 252},
+  {202, 138, 166, 233, 166, 255, 233, 252},
+  {203, 138, 166, 233, 166, 255, 233, 252},
+  {204, 139, 166, 234, 167, 255, 234, 252},
+  {205, 139, 166, 234, 167, 255, 234, 252},
+  {206, 140, 167, 235, 168, 255, 235, 252},
+  {206, 140, 167, 235, 168, 255, 235, 252},
+  {207, 141, 168, 236, 169, 255, 235, 252},
+  {208, 141, 168, 236, 170, 255, 235, 252},
+  {209, 142, 169, 237, 171, 255, 236, 252},
+  {209, 143, 169, 237, 171, 255, 236, 252},
+  {210, 144, 169, 237, 172, 255, 236, 252},
+  {211, 144, 169, 237, 172, 255, 236, 252},
+  {212, 145, 170, 238, 173, 255, 237, 252},
+  {213, 145, 170, 238, 173, 255, 237, 252},
+  {214, 146, 171, 239, 174, 255, 237, 253},
+  {214, 146, 171, 239, 174, 255, 237, 253},
+  {215, 147, 172, 240, 175, 255, 238, 253},
+  {215, 147, 172, 240, 175, 255, 238, 253},
+  {216, 148, 173, 240, 176, 255, 238, 253},
+  {217, 148, 173, 240, 176, 255, 238, 253},
+  {218, 149, 173, 241, 177, 255, 239, 253},
+  {218, 149, 173, 241, 178, 255, 239, 253},
+  {219, 150, 174, 241, 179, 255, 239, 253},
+  {219, 151, 174, 241, 179, 255, 239, 253},
+  {220, 152, 175, 242, 180, 255, 240, 253},
+  {221, 152, 175, 242, 180, 255, 240, 253},
+  {222, 153, 176, 242, 181, 255, 240, 253},
+  {222, 153, 176, 242, 181, 255, 240, 253},
+  {223, 154, 177, 243, 182, 255, 240, 253},
+  {223, 154, 177, 243, 182, 255, 240, 253},
+  {224, 155, 178, 244, 183, 255, 241, 253},
+  {224, 155, 178, 244, 183, 255, 241, 253},
+  {225, 156, 178, 244, 184, 255, 241, 253},
+  {225, 157, 178, 244, 184, 255, 241, 253},
+  {226, 158, 179, 244, 185, 255, 242, 253},
+  {227, 158, 179, 244, 185, 255, 242, 253},
+  {228, 159, 180, 245, 186, 255, 242, 253},
+  {228, 159, 180, 245, 186, 255, 242, 253},
+  {229, 160, 181, 245, 187, 255, 242, 253},
+  {229, 160, 181, 245, 187, 255, 242, 253},
+  {230, 161, 182, 246, 188, 255, 243, 253},
+  {230, 162, 182, 246, 188, 255, 243, 253},
+  {231, 163, 183, 246, 189, 255, 243, 253},
+  {231, 163, 183, 246, 189, 255, 243, 253},
+  {232, 164, 184, 247, 190, 255, 243, 253},
+  {232, 164, 184, 247, 190, 255, 243, 253},
+  {233, 165, 185, 247, 191, 255, 244, 253},
+  {233, 165, 185, 247, 191, 255, 244, 253},
+  {234, 166, 185, 247, 192, 255, 244, 253},
+  {234, 167, 185, 247, 192, 255, 244, 253},
+  {235, 168, 186, 248, 193, 255, 244, 253},
+  {235, 168, 186, 248, 193, 255, 244, 253},
+  {236, 169, 187, 248, 194, 255, 244, 253},
+  {236, 169, 187, 248, 194, 255, 244, 253},
+  {236, 170, 188, 248, 195, 255, 245, 253},
+  {236, 170, 188, 248, 195, 255, 245, 253},
+  {237, 171, 189, 249, 196, 255, 245, 254},
+  {237, 172, 189, 249, 196, 255, 245, 254},
+  {238, 173, 190, 249, 197, 255, 245, 254},
+  {238, 173, 190, 249, 197, 255, 245, 254},
+  {239, 174, 191, 249, 198, 255, 245, 254},
+  {239, 174, 191, 249, 198, 255, 245, 254},
+  {240, 175, 192, 249, 199, 255, 246, 254},
+  {240, 176, 192, 249, 199, 255, 246, 254},
+  {240, 177, 193, 250, 200, 255, 246, 254},
+  {240, 177, 193, 250, 200, 255, 246, 254},
+  {241, 178, 194, 250, 201, 255, 246, 254},
+  {241, 178, 194, 250, 201, 255, 246, 254},
+  {242, 179, 195, 250, 202, 255, 246, 254},
+  {242, 180, 195, 250, 202, 255, 246, 254},
+  {242, 181, 196, 250, 203, 255, 247, 254},
+  {242, 181, 196, 250, 203, 255, 247, 254},
+  {243, 182, 197, 251, 204, 255, 247, 254},
+  {243, 183, 197, 251, 204, 255, 247, 254},
+  {244, 184, 198, 251, 205, 255, 247, 254},
+  {244, 184, 198, 251, 205, 255, 247, 254},
+  {244, 185, 199, 251, 206, 255, 247, 254},
+  {244, 185, 199, 251, 206, 255, 247, 254},
+  {245, 186, 200, 251, 207, 255, 247, 254},
+  {245, 187, 200, 251, 207, 255, 247, 254},
+  {246, 188, 201, 252, 207, 255, 248, 254},
+  {246, 188, 201, 252, 207, 255, 248, 254},
+  {246, 189, 202, 252, 208, 255, 248, 254},
+  {246, 190, 202, 252, 208, 255, 248, 254},
+  {247, 191, 203, 252, 209, 255, 248, 254},
+  {247, 191, 203, 252, 209, 255, 248, 254},
+  {247, 192, 204, 252, 210, 255, 248, 254},
+  {247, 193, 204, 252, 210, 255, 248, 254},
+  {248, 194, 205, 252, 211, 255, 248, 254},
+  {248, 194, 205, 252, 211, 255, 248, 254},
+  {248, 195, 206, 252, 212, 255, 249, 254},
+  {248, 196, 206, 252, 212, 255, 249, 254},
+  {249, 197, 207, 253, 213, 255, 249, 254},
+  {249, 197, 207, 253, 213, 255, 249, 254},
+  {249, 198, 208, 253, 214, 255, 249, 254},
+  {249, 199, 209, 253, 214, 255, 249, 254},
+  {250, 200, 210, 253, 215, 255, 249, 254},
+  {250, 200, 210, 253, 215, 255, 249, 254},
+  {250, 201, 211, 253, 215, 255, 249, 254},
+  {250, 202, 211, 253, 215, 255, 249, 254},
+  {250, 203, 212, 253, 216, 255, 249, 254},
+  {250, 203, 212, 253, 216, 255, 249, 254},
+  {251, 204, 213, 253, 217, 255, 250, 254},
+  {251, 205, 213, 253, 217, 255, 250, 254},
+  {251, 206, 214, 254, 218, 255, 250, 254},
+  {251, 206, 215, 254, 218, 255, 250, 254},
+  {252, 207, 216, 254, 219, 255, 250, 254},
+  {252, 208, 216, 254, 219, 255, 250, 254},
+  {252, 209, 217, 254, 220, 255, 250, 254},
+  {252, 210, 217, 254, 220, 255, 250, 254},
+  {252, 211, 218, 254, 221, 255, 250, 254},
+  {252, 212, 218, 254, 221, 255, 250, 254},
+  {253, 213, 219, 254, 222, 255, 250, 254},
+  {253, 213, 220, 254, 222, 255, 250, 254},
+  {253, 214, 221, 254, 223, 255, 250, 254},
+  {253, 215, 221, 254, 223, 255, 250, 254},
+  {253, 216, 222, 254, 224, 255, 251, 254},
+  {253, 217, 223, 254, 224, 255, 251, 254},
+  {253, 218, 224, 254, 225, 255, 251, 254},
+  {253, 219, 224, 254, 225, 255, 251, 254},
+  {254, 220, 225, 254, 225, 255, 251, 254},
+  {254, 221, 226, 254, 225, 255, 251, 254},
+  {254, 222, 227, 255, 226, 255, 251, 254},
+  {254, 223, 227, 255, 226, 255, 251, 254},
+  {254, 224, 228, 255, 227, 255, 251, 254},
+  {254, 225, 229, 255, 227, 255, 251, 254},
+  {254, 226, 230, 255, 228, 255, 251, 254},
+  {254, 227, 230, 255, 229, 255, 251, 254},
+  {255, 228, 231, 255, 230, 255, 251, 254},
+  {255, 229, 232, 255, 230, 255, 251, 254},
+  {255, 230, 233, 255, 231, 255, 252, 254},
+  {255, 231, 234, 255, 231, 255, 252, 254},
+  {255, 232, 235, 255, 232, 255, 252, 254},
+  {255, 233, 236, 255, 232, 255, 252, 254},
+  {255, 235, 237, 255, 233, 255, 252, 254},
+  {255, 236, 238, 255, 234, 255, 252, 254},
+  {255, 238, 240, 255, 235, 255, 252, 255},
+  {255, 239, 241, 255, 235, 255, 252, 254},
+  {255, 241, 243, 255, 236, 255, 252, 254},
+  {255, 243, 245, 255, 237, 255, 252, 254},
+  {255, 246, 247, 255, 239, 255, 253, 255},
+  {255, 246, 247, 255, 239, 255, 253, 255},
+};
+
+static const vp9_coeff_probs_model default_coef_probs_4x4[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 195,  29, 183 }, {  84,  49, 136 }, {   8,  42,  71 }
+      }, {  // Band 1
+        {  31, 107, 169 }, {  35,  99, 159 }, {  17,  82, 140 },
+        {   8,  66, 114 }, {   2,  44,  76 }, {   1,  19,  32 }
+      }, {  // Band 2
+        {  40, 132, 201 }, {  29, 114, 187 }, {  13,  91, 157 },
+        {   7,  75, 127 }, {   3,  58,  95 }, {   1,  28,  47 }
+      }, {  // Band 3
+        {  69, 142, 221 }, {  42, 122, 201 }, {  15,  91, 159 },
+        {   6,  67, 121 }, {   1,  42,  77 }, {   1,  17,  31 }
+      }, {  // Band 4
+        { 102, 148, 228 }, {  67, 117, 204 }, {  17,  82, 154 },
+        {   6,  59, 114 }, {   2,  39,  75 }, {   1,  15,  29 }
+      }, {  // Band 5
+        { 156,  57, 233 }, { 119,  57, 212 }, {  58,  48, 163 },
+        {  29,  40, 124 }, {  12,  30,  81 }, {   3,  12,  31 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 191, 107, 226 }, { 124, 117, 204 }, {  25,  99, 155 }
+      }, {  // Band 1
+        {  29, 148, 210 }, {  37, 126, 194 }, {   8,  93, 157 },
+        {   2,  68, 118 }, {   1,  39,  69 }, {   1,  17,  33 }
+      }, {  // Band 2
+        {  41, 151, 213 }, {  27, 123, 193 }, {   3,  82, 144 },
+        {   1,  58, 105 }, {   1,  32,  60 }, {   1,  13,  26 }
+      }, {  // Band 3
+        {  59, 159, 220 }, {  23, 126, 198 }, {   4,  88, 151 },
+        {   1,  66, 114 }, {   1,  38,  71 }, {   1,  18,  34 }
+      }, {  // Band 4
+        { 114, 136, 232 }, {  51, 114, 207 }, {  11,  83, 155 },
+        {   3,  56, 105 }, {   1,  33,  65 }, {   1,  17,  34 }
+      }, {  // Band 5
+        { 149,  65, 234 }, { 121,  57, 215 }, {  61,  49, 166 },
+        {  28,  36, 114 }, {  12,  25,  76 }, {   3,  16,  42 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 214,  49, 220 }, { 132,  63, 188 }, {  42,  65, 137 }
+      }, {  // Band 1
+        {  85, 137, 221 }, { 104, 131, 216 }, {  49, 111, 192 },
+        {  21,  87, 155 }, {   2,  49,  87 }, {   1,  16,  28 }
+      }, {  // Band 2
+        {  89, 163, 230 }, {  90, 137, 220 }, {  29, 100, 183 },
+        {  10,  70, 135 }, {   2,  42,  81 }, {   1,  17,  33 }
+      }, {  // Band 3
+        { 108, 167, 237 }, {  55, 133, 222 }, {  15,  97, 179 },
+        {   4,  72, 135 }, {   1,  45,  85 }, {   1,  19,  38 }
+      }, {  // Band 4
+        { 124, 146, 240 }, {  66, 124, 224 }, {  17,  88, 175 },
+        {   4,  58, 122 }, {   1,  36,  75 }, {   1,  18,  37 }
+      }, {  //  Band 5
+        { 141,  79, 241 }, { 126,  70, 227 }, {  66,  58, 182 },
+        {  30,  44, 136 }, {  12,  34,  96 }, {   2,  20,  47 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 229,  99, 249 }, { 143, 111, 235 }, {  46, 109, 192 }
+      }, {  // Band 1
+        {  82, 158, 236 }, {  94, 146, 224 }, {  25, 117, 191 },
+        {   9,  87, 149 }, {   3,  56,  99 }, {   1,  33,  57 }
+      }, {  // Band 2
+        {  83, 167, 237 }, {  68, 145, 222 }, {  10, 103, 177 },
+        {   2,  72, 131 }, {   1,  41,  79 }, {   1,  20,  39 }
+      }, {  // Band 3
+        {  99, 167, 239 }, {  47, 141, 224 }, {  10, 104, 178 },
+        {   2,  73, 133 }, {   1,  44,  85 }, {   1,  22,  47 }
+      }, {  // Band 4
+        { 127, 145, 243 }, {  71, 129, 228 }, {  17,  93, 177 },
+        {   3,  61, 124 }, {   1,  41,  84 }, {   1,  21,  52 }
+      }, {  // Band 5
+        { 157,  78, 244 }, { 140,  72, 231 }, {  69,  58, 184 },
+        {  31,  44, 137 }, {  14,  38, 105 }, {   8,  23,  61 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_8x8[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 125,  34, 187 }, {  52,  41, 133 }, {   6,  31,  56 }
+      }, {  // Band 1
+        {  37, 109, 153 }, {  51, 102, 147 }, {  23,  87, 128 },
+        {   8,  67, 101 }, {   1,  41,  63 }, {   1,  19,  29 }
+      }, {  // Band 2
+        {  31, 154, 185 }, {  17, 127, 175 }, {   6,  96, 145 },
+        {   2,  73, 114 }, {   1,  51,  82 }, {   1,  28,  45 }
+      }, {  // Band 3
+        {  23, 163, 200 }, {  10, 131, 185 }, {   2,  93, 148 },
+        {   1,  67, 111 }, {   1,  41,  69 }, {   1,  14,  24 }
+      }, {  // Band 4
+        {  29, 176, 217 }, {  12, 145, 201 }, {   3, 101, 156 },
+        {   1,  69, 111 }, {   1,  39,  63 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  57, 192, 233 }, {  25, 154, 215 }, {   6, 109, 167 },
+        {   3,  78, 118 }, {   1,  48,  69 }, {   1,  21,  29 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 202, 105, 245 }, { 108, 106, 216 }, {  18,  90, 144 }
+      }, {  // Band 1
+        {  33, 172, 219 }, {  64, 149, 206 }, {  14, 117, 177 },
+        {   5,  90, 141 }, {   2,  61,  95 }, {   1,  37,  57 }
+      }, {  // Band 2
+        {  33, 179, 220 }, {  11, 140, 198 }, {   1,  89, 148 },
+        {   1,  60, 104 }, {   1,  33,  57 }, {   1,  12,  21 }
+      }, {  // Band 3
+        {  30, 181, 221 }, {   8, 141, 198 }, {   1,  87, 145 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  32, 186, 224 }, {   7, 142, 198 }, {   1,  86, 143 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  22 }
+      }, {  // Band 5
+        {  57, 192, 227 }, {  20, 143, 204 }, {   3,  96, 154 },
+        {   1,  68, 112 }, {   1,  42,  69 }, {   1,  19,  32 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 212,  35, 215 }, { 113,  47, 169 }, {  29,  48, 105 }
+      }, {  // Band 1
+        {  74, 129, 203 }, { 106, 120, 203 }, {  49, 107, 178 },
+        {  19,  84, 144 }, {   4,  50,  84 }, {   1,  15,  25 }
+      }, {  // Band 2
+        {  71, 172, 217 }, {  44, 141, 209 }, {  15, 102, 173 },
+        {   6,  76, 133 }, {   2,  51,  89 }, {   1,  24,  42 }
+      }, {  // Band 3
+        {  64, 185, 231 }, {  31, 148, 216 }, {   8, 103, 175 },
+        {   3,  74, 131 }, {   1,  46,  81 }, {   1,  18,  30 }
+      }, {  // Band 4
+        {  65, 196, 235 }, {  25, 157, 221 }, {   5, 105, 174 },
+        {   1,  67, 120 }, {   1,  38,  69 }, {   1,  15,  30 }
+      }, {  // Band 5
+        {  65, 204, 238 }, {  30, 156, 224 }, {   7, 107, 177 },
+        {   2,  70, 124 }, {   1,  42,  73 }, {   1,  18,  34 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 225,  86, 251 }, { 144, 104, 235 }, {  42,  99, 181 }
+      }, {  // Band 1
+        {  85, 175, 239 }, { 112, 165, 229 }, {  29, 136, 200 },
+        {  12, 103, 162 }, {   6,  77, 123 }, {   2,  53,  84 }
+      }, {  // Band 2
+        {  75, 183, 239 }, {  30, 155, 221 }, {   3, 106, 171 },
+        {   1,  74, 128 }, {   1,  44,  76 }, {   1,  17,  28 }
+      }, {  // Band 3
+        {  73, 185, 240 }, {  27, 159, 222 }, {   2, 107, 172 },
+        {   1,  75, 127 }, {   1,  42,  73 }, {   1,  17,  29 }
+      }, {  // Band 4
+        {  62, 190, 238 }, {  21, 159, 222 }, {   2, 107, 172 },
+        {   1,  72, 122 }, {   1,  40,  71 }, {   1,  18,  32 }
+      }, {  // Band 5
+        {  61, 199, 240 }, {  27, 161, 226 }, {   4, 113, 180 },
+        {   1,  76, 129 }, {   1,  46,  80 }, {   1,  23,  41 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_16x16[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {   7,  27, 153 }, {   5,  30,  95 }, {   1,  16,  30 }
+      }, {  // Band 1
+        {  50,  75, 127 }, {  57,  75, 124 }, {  27,  67, 108 },
+        {  10,  54,  86 }, {   1,  33,  52 }, {   1,  12,  18 }
+      }, {  // Band 2
+        {  43, 125, 151 }, {  26, 108, 148 }, {   7,  83, 122 },
+        {   2,  59,  89 }, {   1,  38,  60 }, {   1,  17,  27 }
+      }, {  // Band 3
+        {  23, 144, 163 }, {  13, 112, 154 }, {   2,  75, 117 },
+        {   1,  50,  81 }, {   1,  31,  51 }, {   1,  14,  23 }
+      }, {  // Band 4
+        {  18, 162, 185 }, {   6, 123, 171 }, {   1,  78, 125 },
+        {   1,  51,  86 }, {   1,  31,  54 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  15, 199, 227 }, {   3, 150, 204 }, {   1,  91, 146 },
+        {   1,  55,  95 }, {   1,  30,  53 }, {   1,  11,  20 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  19,  55, 240 }, {  19,  59, 196 }, {   3,  52, 105 }
+      }, {  // Band 1
+        {  41, 166, 207 }, { 104, 153, 199 }, {  31, 123, 181 },
+        {  14, 101, 152 }, {   5,  72, 106 }, {   1,  36,  52 }
+      }, {  // Band 2
+        {  35, 176, 211 }, {  12, 131, 190 }, {   2,  88, 144 },
+        {   1,  60, 101 }, {   1,  36,  60 }, {   1,  16,  28 }
+      }, {  // Band 3
+        {  28, 183, 213 }, {   8, 134, 191 }, {   1,  86, 142 },
+        {   1,  56,  96 }, {   1,  30,  53 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  20, 190, 215 }, {   4, 135, 192 }, {   1,  84, 139 },
+        {   1,  53,  91 }, {   1,  28,  49 }, {   1,  11,  20 }
+      }, {  // Band 5
+        {  13, 196, 216 }, {   2, 137, 192 }, {   1,  86, 143 },
+        {   1,  57,  99 }, {   1,  32,  56 }, {   1,  13,  24 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 211,  29, 217 }, {  96,  47, 156 }, {  22,  43,  87 }
+      }, {  // Band 1
+        {  78, 120, 193 }, { 111, 116, 186 }, {  46, 102, 164 },
+        {  15,  80, 128 }, {   2,  49,  76 }, {   1,  18,  28 }
+      }, {  // Band 2
+        {  71, 161, 203 }, {  42, 132, 192 }, {  10,  98, 150 },
+        {   3,  69, 109 }, {   1,  44,  70 }, {   1,  18,  29 }
+      }, {  // Band 3
+        {  57, 186, 211 }, {  30, 140, 196 }, {   4,  93, 146 },
+        {   1,  62, 102 }, {   1,  38,  65 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  47, 199, 217 }, {  14, 145, 196 }, {   1,  88, 142 },
+        {   1,  57,  98 }, {   1,  36,  62 }, {   1,  15,  26 }
+      }, {  // Band 5
+        {  26, 219, 229 }, {   5, 155, 207 }, {   1,  94, 151 },
+        {   1,  60, 104 }, {   1,  36,  62 }, {   1,  16,  28 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 233,  29, 248 }, { 146,  47, 220 }, {  43,  52, 140 }
+      }, {  // Band 1
+        { 100, 163, 232 }, { 179, 161, 222 }, {  63, 142, 204 },
+        {  37, 113, 174 }, {  26,  89, 137 }, {  18,  68,  97 }
+      }, {  // Band 2
+        {  85, 181, 230 }, {  32, 146, 209 }, {   7, 100, 164 },
+        {   3,  71, 121 }, {   1,  45,  77 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  65, 187, 230 }, {  20, 148, 207 }, {   2,  97, 159 },
+        {   1,  68, 116 }, {   1,  40,  70 }, {   1,  14,  29 }
+      }, {  // Band 4
+        {  40, 194, 227 }, {   8, 147, 204 }, {   1,  94, 155 },
+        {   1,  65, 112 }, {   1,  39,  66 }, {   1,  14,  26 }
+      }, {  // Band 5
+        {  16, 208, 228 }, {   3, 151, 207 }, {   1,  98, 160 },
+        {   1,  67, 117 }, {   1,  41,  74 }, {   1,  17,  31 }
+      }
+    }
+  }
+};
+
+static const vp9_coeff_probs_model default_coef_probs_32x32[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {  17,  38, 140 }, {   7,  34,  80 }, {   1,  17,  29 }
+      }, {  // Band 1
+        {  37,  75, 128 }, {  41,  76, 128 }, {  26,  66, 116 },
+        {  12,  52,  94 }, {   2,  32,  55 }, {   1,  10,  16 }
+      }, {  // Band 2
+        {  50, 127, 154 }, {  37, 109, 152 }, {  16,  82, 121 },
+        {   5,  59,  85 }, {   1,  35,  54 }, {   1,  13,  20 }
+      }, {  // Band 3
+        {  40, 142, 167 }, {  17, 110, 157 }, {   2,  71, 112 },
+        {   1,  44,  72 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }, {  // Band 4
+        {  30, 175, 188 }, {   9, 124, 169 }, {   1,  74, 116 },
+        {   1,  48,  78 }, {   1,  30,  49 }, {   1,  11,  18 }
+      }, {  // Band 5
+        {  10, 222, 223 }, {   2, 150, 194 }, {   1,  83, 128 },
+        {   1,  48,  79 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  36,  41, 235 }, {  29,  36, 193 }, {  10,  27, 111 }
+      }, {  // Band 1
+        {  85, 165, 222 }, { 177, 162, 215 }, { 110, 135, 195 },
+        {  57, 113, 168 }, {  23,  83, 120 }, {  10,  49,  61 }
+      }, {  // Band 2
+        {  85, 190, 223 }, {  36, 139, 200 }, {   5,  90, 146 },
+        {   1,  60, 103 }, {   1,  38,  65 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  72, 202, 223 }, {  23, 141, 199 }, {   2,  86, 140 },
+        {   1,  56,  97 }, {   1,  36,  61 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  55, 218, 225 }, {  13, 145, 200 }, {   1,  86, 141 },
+        {   1,  57,  99 }, {   1,  35,  61 }, {   1,  13,  22 }
+      }, {  // Band 5
+        {  15, 235, 212 }, {   1, 132, 184 }, {   1,  84, 139 },
+        {   1,  57,  97 }, {   1,  34,  56 }, {   1,  14,  23 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 181,  21, 201 }, {  61,  37, 123 }, {  10,  38,  71 }
+      }, {  // Band 1
+        {  47, 106, 172 }, {  95, 104, 173 }, {  42,  93, 159 },
+        {  18,  77, 131 }, {   4,  50,  81 }, {   1,  17,  23 }
+      }, {  // Band 2
+        {  62, 147, 199 }, {  44, 130, 189 }, {  28, 102, 154 },
+        {  18,  75, 115 }, {   2,  44,  65 }, {   1,  12,  19 }
+      }, {  // Band 3
+        {  55, 153, 210 }, {  24, 130, 194 }, {   3,  93, 146 },
+        {   1,  61,  97 }, {   1,  31,  50 }, {   1,  10,  16 }
+      }, {  // Band 4
+        {  49, 186, 223 }, {  17, 148, 204 }, {   1,  96, 142 },
+        {   1,  53,  83 }, {   1,  26,  44 }, {   1,  11,  17 }
+      }, {  // Band 5
+        {  13, 217, 212 }, {   2, 136, 180 }, {   1,  78, 124 },
+        {   1,  50,  83 }, {   1,  29,  49 }, {   1,  14,  23 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 197,  13, 247 }, {  82,  17, 222 }, {  25,  17, 162 }
+      }, {  // Band 1
+        { 126, 186, 247 }, { 234, 191, 243 }, { 176, 177, 234 },
+        { 104, 158, 220 }, {  66, 128, 186 }, {  55,  90, 137 }
+      }, {  // Band 2
+        { 111, 197, 242 }, {  46, 158, 219 }, {   9, 104, 171 },
+        {   2,  65, 125 }, {   1,  44,  80 }, {   1,  17,  91 }
+      }, {  // Band 3
+        { 104, 208, 245 }, {  39, 168, 224 }, {   3, 109, 162 },
+        {   1,  79, 124 }, {   1,  50, 102 }, {   1,  43, 102 }
+      }, {  // Band 4
+        {  84, 220, 246 }, {  31, 177, 231 }, {   2, 115, 180 },
+        {   1,  79, 134 }, {   1,  55,  77 }, {   1,  60,  79 }
+      }, {  // Band 5
+        {  43, 243, 240 }, {   8, 180, 217 }, {   1, 115, 166 },
+        {   1,  84, 121 }, {   1,  51,  67 }, {   1,  16,   6 }
+      }
+    }
+  }
+};
+
+static void extend_to_full_distribution(vp9_prob *probs, vp9_prob p) {
+  memcpy(probs, vp9_pareto8_full[p = 0 ? 0 : p - 1],
+         MODEL_NODES * sizeof(vp9_prob));
+}
+
+void vp9_model_to_full_probs(const vp9_prob *model, vp9_prob *full) {
+  if (full != model)
+    memcpy(full, model, sizeof(vp9_prob) * UNCONSTRAINED_NODES);
+  extend_to_full_distribution(&full[UNCONSTRAINED_NODES], model[PIVOT_NODE]);
+}
+
+void vp9_default_coef_probs(VP9_COMMON *cm) {
+  vp9_copy(cm->fc->coef_probs[TX_4X4], default_coef_probs_4x4);
+  vp9_copy(cm->fc->coef_probs[TX_8X8], default_coef_probs_8x8);
+  vp9_copy(cm->fc->coef_probs[TX_16X16], default_coef_probs_16x16);
+  vp9_copy(cm->fc->coef_probs[TX_32X32], default_coef_probs_32x32);
+}
+
+#define COEF_COUNT_SAT 24
+#define COEF_MAX_UPDATE_FACTOR 112
+#define COEF_COUNT_SAT_KEY 24
+#define COEF_MAX_UPDATE_FACTOR_KEY 112
+#define COEF_COUNT_SAT_AFTER_KEY 24
+#define COEF_MAX_UPDATE_FACTOR_AFTER_KEY 128
+
+static void adapt_coef_probs(VP9_COMMON *cm, TX_SIZE tx_size,
+                             unsigned int count_sat,
+                             unsigned int update_factor) {
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  vp9_coeff_probs_model *const probs = cm->fc->coef_probs[tx_size];
+  const vp9_coeff_probs_model *const pre_probs = pre_fc->coef_probs[tx_size];
+  vp9_coeff_count_model *counts = cm->counts.coef[tx_size];
+  unsigned int (*eob_counts)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
+      cm->counts.eob_branch[tx_size];
+  int i, j, k, l, m;
+
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+          const int n0 = counts[i][j][k][l][ZERO_TOKEN];
+          const int n1 = counts[i][j][k][l][ONE_TOKEN];
+          const int n2 = counts[i][j][k][l][TWO_TOKEN];
+          const int neob = counts[i][j][k][l][EOB_MODEL_TOKEN];
+          const unsigned int branch_ct[UNCONSTRAINED_NODES][2] = {
+            { neob, eob_counts[i][j][k][l] - neob },
+            { n0, n1 + n2 },
+            { n1, n2 }
+          };
+          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+            probs[i][j][k][l][m] = merge_probs(pre_probs[i][j][k][l][m],
+                                               branch_ct[m],
+                                               count_sat, update_factor);
+        }
+}
+
+void vp9_adapt_coef_probs(VP9_COMMON *cm) {
+  TX_SIZE t;
+  unsigned int count_sat, update_factor;
+
+  if (frame_is_intra_only(cm)) {
+    update_factor = COEF_MAX_UPDATE_FACTOR_KEY;
+    count_sat = COEF_COUNT_SAT_KEY;
+  } else if (cm->last_frame_type == KEY_FRAME) {
+    update_factor = COEF_MAX_UPDATE_FACTOR_AFTER_KEY;  /* adapt quickly */
+    count_sat = COEF_COUNT_SAT_AFTER_KEY;
+  } else {
+    update_factor = COEF_MAX_UPDATE_FACTOR;
+    count_sat = COEF_COUNT_SAT;
+  }
+  for (t = TX_4X4; t <= TX_32X32; t++)
+    adapt_coef_probs(cm, t, count_sat, update_factor);
+}
diff --git a/media/libvpx/vp9/common/vp9_entropy.h b/media/libvpx/vp9/common/vp9_entropy.h
new file mode 100644
index 000000000..4e02630e6
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropy.h
@@ -0,0 +1,216 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENTROPY_H_
+#define VP9_COMMON_VP9_ENTROPY_H_
+
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DIFF_UPDATE_PROB 252
+
+// Coefficient token alphabet
+#define ZERO_TOKEN      0   // 0     Extra Bits 0+0
+#define ONE_TOKEN       1   // 1     Extra Bits 0+1
+#define TWO_TOKEN       2   // 2     Extra Bits 0+1
+#define THREE_TOKEN     3   // 3     Extra Bits 0+1
+#define FOUR_TOKEN      4   // 4     Extra Bits 0+1
+#define CATEGORY1_TOKEN 5   // 5-6   Extra Bits 1+1
+#define CATEGORY2_TOKEN 6   // 7-10  Extra Bits 2+1
+#define CATEGORY3_TOKEN 7   // 11-18 Extra Bits 3+1
+#define CATEGORY4_TOKEN 8   // 19-34 Extra Bits 4+1
+#define CATEGORY5_TOKEN 9   // 35-66 Extra Bits 5+1
+#define CATEGORY6_TOKEN 10  // 67+   Extra Bits 14+1
+#define EOB_TOKEN       11  // EOB   Extra Bits 0+0
+
+#define ENTROPY_TOKENS 12
+
+#define ENTROPY_NODES 11
+
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_pt_energy_class[ENTROPY_TOKENS]);
+
+#define CAT1_MIN_VAL    5
+#define CAT2_MIN_VAL    7
+#define CAT3_MIN_VAL   11
+#define CAT4_MIN_VAL   19
+#define CAT5_MIN_VAL   35
+#define CAT6_MIN_VAL   67
+
+// Extra bit probabilities.
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob[14]);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob_high10[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob_high10[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob_high10[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob_high10[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob_high10[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob_high10[16]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat1_prob_high12[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat2_prob_high12[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat3_prob_high12[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat4_prob_high12[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat5_prob_high12[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_cat6_prob_high12[18]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#define EOB_MODEL_TOKEN 3
+extern const vp9_tree_index vp9_coefmodel_tree[];
+
+typedef struct {
+  const vp9_tree_index *tree;
+  const vp9_prob *prob;
+  int len;
+  int base_val;
+  const int16_t *cost;
+} vp9_extra_bit;
+
+// indexed by token value
+extern const vp9_extra_bit vp9_extra_bits[ENTROPY_TOKENS];
+#if CONFIG_VP9_HIGHBITDEPTH
+extern const vp9_extra_bit vp9_extra_bits_high10[ENTROPY_TOKENS];
+extern const vp9_extra_bit vp9_extra_bits_high12[ENTROPY_TOKENS];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#define DCT_MAX_VALUE           16384
+#if CONFIG_VP9_HIGHBITDEPTH
+#define DCT_MAX_VALUE_HIGH10    65536
+#define DCT_MAX_VALUE_HIGH12   262144
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+/* Coefficients are predicted via a 3-dimensional probability table. */
+
+#define REF_TYPES 2  // intra=0, inter=1
+
+/* Middle dimension reflects the coefficient position within the transform. */
+#define COEF_BANDS 6
+
+/* Inside dimension is measure of nearby complexity, that reflects the energy
+   of nearby coefficients are nonzero.  For the first coefficient (DC, unless
+   block type is 0), we look at the (already encoded) blocks above and to the
+   left of the current block.  The context index is then the number (0,1,or 2)
+   of these blocks having nonzero coefficients.
+   After decoding a coefficient, the measure is determined by the size of the
+   most recently decoded coefficient.
+   Note that the intuitive meaning of this measure changes as coefficients
+   are decoded, e.g., prior to the first token, a zero means that my neighbors
+   are empty while, after the first token, because of the use of end-of-block,
+   a zero means we just decoded a zero and hence guarantees that a non-zero
+   coefficient will appear later in this block.  However, this shift
+   in meaning is perfectly OK because our context depends also on the
+   coefficient band (and since zigzag positions 0, 1, and 2 are in
+   distinct bands). */
+
+#define COEFF_CONTEXTS 6
+#define BAND_COEFF_CONTEXTS(band) ((band) == 0 ? 3 : COEFF_CONTEXTS)
+
+// #define ENTROPY_STATS
+
+typedef unsigned int vp9_coeff_count[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_TOKENS];
+typedef unsigned int vp9_coeff_stats[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_NODES][2];
+
+#define SUBEXP_PARAM                4   /* Subexponential code parameter */
+#define MODULUS_PARAM               13  /* Modulus parameter */
+
+struct VP9Common;
+void vp9_default_coef_probs(struct VP9Common *cm);
+void vp9_adapt_coef_probs(struct VP9Common *cm);
+
+// This is the index in the scan order beyond which all coefficients for
+// 8x8 transform and above are in the top band.
+// This macro is currently unused but may be used by certain implementations
+#define MAXBAND_INDEX 21
+
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_coefband_trans_8x8plus[1024]);
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_coefband_trans_4x4[16]);
+
+static INLINE const uint8_t *get_band_translate(TX_SIZE tx_size) {
+  return tx_size == TX_4X4 ? vp9_coefband_trans_4x4
+                           : vp9_coefband_trans_8x8plus;
+}
+
+// 128 lists of probabilities are stored for the following ONE node probs:
+// 1, 3, 5, 7, ..., 253, 255
+// In between probabilities are interpolated linearly
+
+#define COEFF_PROB_MODELS 256
+
+#define UNCONSTRAINED_NODES         3
+
+#define PIVOT_NODE                  2   // which node is pivot
+
+#define MODEL_NODES (ENTROPY_NODES - UNCONSTRAINED_NODES)
+extern const vp9_tree_index vp9_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)];
+extern const vp9_prob vp9_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES];
+
+typedef vp9_prob vp9_coeff_probs_model[REF_TYPES][COEF_BANDS]
+                                      [COEFF_CONTEXTS][UNCONSTRAINED_NODES];
+
+typedef unsigned int vp9_coeff_count_model[REF_TYPES][COEF_BANDS]
+                                          [COEFF_CONTEXTS]
+                                          [UNCONSTRAINED_NODES + 1];
+
+void vp9_model_to_full_probs(const vp9_prob *model, vp9_prob *full);
+
+typedef char ENTROPY_CONTEXT;
+
+static INLINE int combine_entropy_contexts(ENTROPY_CONTEXT a,
+                                           ENTROPY_CONTEXT b) {
+  return (a != 0) + (b != 0);
+}
+
+static INLINE int get_entropy_context(TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
+                                      const ENTROPY_CONTEXT *l) {
+  ENTROPY_CONTEXT above_ec = 0, left_ec = 0;
+
+  switch (tx_size) {
+    case TX_4X4:
+      above_ec = a[0] != 0;
+      left_ec = l[0] != 0;
+      break;
+    case TX_8X8:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec  = !!*(const uint16_t *)l;
+      break;
+    case TX_16X16:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec  = !!*(const uint32_t *)l;
+      break;
+    case TX_32X32:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec  = !!*(const uint64_t *)l;
+      break;
+    default:
+      assert(0 && "Invalid transform size.");
+      break;
+  }
+
+  return combine_entropy_contexts(above_ec, left_ec);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPY_H_
diff --git a/media/libvpx/vp9/common/vp9_entropymode.c b/media/libvpx/vp9/common/vp9_entropymode.c
new file mode 100644
index 000000000..424451fee
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropymode.c
@@ -0,0 +1,469 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_seg_common.h"
+
+const vp9_prob vp9_kf_y_mode_prob[INTRA_MODES][INTRA_MODES][INTRA_MODES - 1] = {
+  {  // above = dc
+    { 137,  30,  42, 148, 151, 207,  70,  52,  91 },  // left = dc
+    {  92,  45, 102, 136, 116, 180,  74,  90, 100 },  // left = v
+    {  73,  32,  19, 187, 222, 215,  46,  34, 100 },  // left = h
+    {  91,  30,  32, 116, 121, 186,  93,  86,  94 },  // left = d45
+    {  72,  35,  36, 149,  68, 206,  68,  63, 105 },  // left = d135
+    {  73,  31,  28, 138,  57, 124,  55, 122, 151 },  // left = d117
+    {  67,  23,  21, 140, 126, 197,  40,  37, 171 },  // left = d153
+    {  86,  27,  28, 128, 154, 212,  45,  43,  53 },  // left = d207
+    {  74,  32,  27, 107,  86, 160,  63, 134, 102 },  // left = d63
+    {  59,  67,  44, 140, 161, 202,  78,  67, 119 }   // left = tm
+  }, {  // above = v
+    {  63,  36, 126, 146, 123, 158,  60,  90,  96 },  // left = dc
+    {  43,  46, 168, 134, 107, 128,  69, 142,  92 },  // left = v
+    {  44,  29,  68, 159, 201, 177,  50,  57,  77 },  // left = h
+    {  58,  38,  76, 114,  97, 172,  78, 133,  92 },  // left = d45
+    {  46,  41,  76, 140,  63, 184,  69, 112,  57 },  // left = d135
+    {  38,  32,  85, 140,  46, 112,  54, 151, 133 },  // left = d117
+    {  39,  27,  61, 131, 110, 175,  44,  75, 136 },  // left = d153
+    {  52,  30,  74, 113, 130, 175,  51,  64,  58 },  // left = d207
+    {  47,  35,  80, 100,  74, 143,  64, 163,  74 },  // left = d63
+    {  36,  61, 116, 114, 128, 162,  80, 125,  82 }   // left = tm
+  }, {  // above = h
+    {  82,  26,  26, 171, 208, 204,  44,  32, 105 },  // left = dc
+    {  55,  44,  68, 166, 179, 192,  57,  57, 108 },  // left = v
+    {  42,  26,  11, 199, 241, 228,  23,  15,  85 },  // left = h
+    {  68,  42,  19, 131, 160, 199,  55,  52,  83 },  // left = d45
+    {  58,  50,  25, 139, 115, 232,  39,  52, 118 },  // left = d135
+    {  50,  35,  33, 153, 104, 162,  64,  59, 131 },  // left = d117
+    {  44,  24,  16, 150, 177, 202,  33,  19, 156 },  // left = d153
+    {  55,  27,  12, 153, 203, 218,  26,  27,  49 },  // left = d207
+    {  53,  49,  21, 110, 116, 168,  59,  80,  76 },  // left = d63
+    {  38,  72,  19, 168, 203, 212,  50,  50, 107 }   // left = tm
+  }, {  // above = d45
+    { 103,  26,  36, 129, 132, 201,  83,  80,  93 },  // left = dc
+    {  59,  38,  83, 112, 103, 162,  98, 136,  90 },  // left = v
+    {  62,  30,  23, 158, 200, 207,  59,  57,  50 },  // left = h
+    {  67,  30,  29,  84,  86, 191, 102,  91,  59 },  // left = d45
+    {  60,  32,  33, 112,  71, 220,  64,  89, 104 },  // left = d135
+    {  53,  26,  34, 130,  56, 149,  84, 120, 103 },  // left = d117
+    {  53,  21,  23, 133, 109, 210,  56,  77, 172 },  // left = d153
+    {  77,  19,  29, 112, 142, 228,  55,  66,  36 },  // left = d207
+    {  61,  29,  29,  93,  97, 165,  83, 175, 162 },  // left = d63
+    {  47,  47,  43, 114, 137, 181, 100,  99,  95 }   // left = tm
+  }, {  // above = d135
+    {  69,  23,  29, 128,  83, 199,  46,  44, 101 },  // left = dc
+    {  53,  40,  55, 139,  69, 183,  61,  80, 110 },  // left = v
+    {  40,  29,  19, 161, 180, 207,  43,  24,  91 },  // left = h
+    {  60,  34,  19, 105,  61, 198,  53,  64,  89 },  // left = d45
+    {  52,  31,  22, 158,  40, 209,  58,  62,  89 },  // left = d135
+    {  44,  31,  29, 147,  46, 158,  56, 102, 198 },  // left = d117
+    {  35,  19,  12, 135,  87, 209,  41,  45, 167 },  // left = d153
+    {  55,  25,  21, 118,  95, 215,  38,  39,  66 },  // left = d207
+    {  51,  38,  25, 113,  58, 164,  70,  93,  97 },  // left = d63
+    {  47,  54,  34, 146, 108, 203,  72, 103, 151 }   // left = tm
+  }, {  // above = d117
+    {  64,  19,  37, 156,  66, 138,  49,  95, 133 },  // left = dc
+    {  46,  27,  80, 150,  55, 124,  55, 121, 135 },  // left = v
+    {  36,  23,  27, 165, 149, 166,  54,  64, 118 },  // left = h
+    {  53,  21,  36, 131,  63, 163,  60, 109,  81 },  // left = d45
+    {  40,  26,  35, 154,  40, 185,  51,  97, 123 },  // left = d135
+    {  35,  19,  34, 179,  19,  97,  48, 129, 124 },  // left = d117
+    {  36,  20,  26, 136,  62, 164,  33,  77, 154 },  // left = d153
+    {  45,  18,  32, 130,  90, 157,  40,  79,  91 },  // left = d207
+    {  45,  26,  28, 129,  45, 129,  49, 147, 123 },  // left = d63
+    {  38,  44,  51, 136,  74, 162,  57,  97, 121 }   // left = tm
+  }, {  // above = d153
+    {  75,  17,  22, 136, 138, 185,  32,  34, 166 },  // left = dc
+    {  56,  39,  58, 133, 117, 173,  48,  53, 187 },  // left = v
+    {  35,  21,  12, 161, 212, 207,  20,  23, 145 },  // left = h
+    {  56,  29,  19, 117, 109, 181,  55,  68, 112 },  // left = d45
+    {  47,  29,  17, 153,  64, 220,  59,  51, 114 },  // left = d135
+    {  46,  16,  24, 136,  76, 147,  41,  64, 172 },  // left = d117
+    {  34,  17,  11, 108, 152, 187,  13,  15, 209 },  // left = d153
+    {  51,  24,  14, 115, 133, 209,  32,  26, 104 },  // left = d207
+    {  55,  30,  18, 122,  79, 179,  44,  88, 116 },  // left = d63
+    {  37,  49,  25, 129, 168, 164,  41,  54, 148 }   // left = tm
+  }, {  // above = d207
+    {  82,  22,  32, 127, 143, 213,  39,  41,  70 },  // left = dc
+    {  62,  44,  61, 123, 105, 189,  48,  57,  64 },  // left = v
+    {  47,  25,  17, 175, 222, 220,  24,  30,  86 },  // left = h
+    {  68,  36,  17, 106, 102, 206,  59,  74,  74 },  // left = d45
+    {  57,  39,  23, 151,  68, 216,  55,  63,  58 },  // left = d135
+    {  49,  30,  35, 141,  70, 168,  82,  40, 115 },  // left = d117
+    {  51,  25,  15, 136, 129, 202,  38,  35, 139 },  // left = d153
+    {  68,  26,  16, 111, 141, 215,  29,  28,  28 },  // left = d207
+    {  59,  39,  19, 114,  75, 180,  77, 104,  42 },  // left = d63
+    {  40,  61,  26, 126, 152, 206,  61,  59,  93 }   // left = tm
+  }, {  // above = d63
+    {  78,  23,  39, 111, 117, 170,  74, 124,  94 },  // left = dc
+    {  48,  34,  86, 101,  92, 146,  78, 179, 134 },  // left = v
+    {  47,  22,  24, 138, 187, 178,  68,  69,  59 },  // left = h
+    {  56,  25,  33, 105, 112, 187,  95, 177, 129 },  // left = d45
+    {  48,  31,  27, 114,  63, 183,  82, 116,  56 },  // left = d135
+    {  43,  28,  37, 121,  63, 123,  61, 192, 169 },  // left = d117
+    {  42,  17,  24, 109,  97, 177,  56,  76, 122 },  // left = d153
+    {  58,  18,  28, 105, 139, 182,  70,  92,  63 },  // left = d207
+    {  46,  23,  32,  74,  86, 150,  67, 183,  88 },  // left = d63
+    {  36,  38,  48,  92, 122, 165,  88, 137,  91 }   // left = tm
+  }, {  // above = tm
+    {  65,  70,  60, 155, 159, 199,  61,  60,  81 },  // left = dc
+    {  44,  78, 115, 132, 119, 173,  71, 112,  93 },  // left = v
+    {  39,  38,  21, 184, 227, 206,  42,  32,  64 },  // left = h
+    {  58,  47,  36, 124, 137, 193,  80,  82,  78 },  // left = d45
+    {  49,  50,  35, 144,  95, 205,  63,  78,  59 },  // left = d135
+    {  41,  53,  52, 148,  71, 142,  65, 128,  51 },  // left = d117
+    {  40,  36,  28, 143, 143, 202,  40,  55, 137 },  // left = d153
+    {  52,  34,  29, 129, 183, 227,  42,  35,  43 },  // left = d207
+    {  42,  44,  44, 104, 105, 164,  64, 130,  80 },  // left = d63
+    {  43,  81,  53, 140, 169, 204,  68,  84,  72 }   // left = tm
+  }
+};
+
+const vp9_prob vp9_kf_uv_mode_prob[INTRA_MODES][INTRA_MODES - 1] = {
+  { 144,  11,  54, 157, 195, 130,  46,  58, 108 },  // y = dc
+  { 118,  15, 123, 148, 131, 101,  44,  93, 131 },  // y = v
+  { 113,  12,  23, 188, 226, 142,  26,  32, 125 },  // y = h
+  { 120,  11,  50, 123, 163, 135,  64,  77, 103 },  // y = d45
+  { 113,   9,  36, 155, 111, 157,  32,  44, 161 },  // y = d135
+  { 116,   9,  55, 176,  76,  96,  37,  61, 149 },  // y = d117
+  { 115,   9,  28, 141, 161, 167,  21,  25, 193 },  // y = d153
+  { 120,  12,  32, 145, 195, 142,  32,  38,  86 },  // y = d207
+  { 116,  12,  64, 120, 140, 125,  49, 115, 121 },  // y = d63
+  { 102,  19,  66, 162, 182, 122,  35,  59, 128 }   // y = tm
+};
+
+static const vp9_prob default_if_y_probs[BLOCK_SIZE_GROUPS][INTRA_MODES - 1] = {
+  {  65,  32,  18, 144, 162, 194,  41,  51,  98 },  // block_size < 8x8
+  { 132,  68,  18, 165, 217, 196,  45,  40,  78 },  // block_size < 16x16
+  { 173,  80,  19, 176, 240, 193,  64,  35,  46 },  // block_size < 32x32
+  { 221, 135,  38, 194, 248, 121,  96,  85,  29 }   // block_size >= 32x32
+};
+
+static const vp9_prob default_if_uv_probs[INTRA_MODES][INTRA_MODES - 1] = {
+  { 120,   7,  76, 176, 208, 126,  28,  54, 103 },  // y = dc
+  {  48,  12, 154, 155, 139,  90,  34, 117, 119 },  // y = v
+  {  67,   6,  25, 204, 243, 158,  13,  21,  96 },  // y = h
+  {  97,   5,  44, 131, 176, 139,  48,  68,  97 },  // y = d45
+  {  83,   5,  42, 156, 111, 152,  26,  49, 152 },  // y = d135
+  {  80,   5,  58, 178,  74,  83,  33,  62, 145 },  // y = d117
+  {  86,   5,  32, 154, 192, 168,  14,  22, 163 },  // y = d153
+  {  85,   5,  32, 156, 216, 148,  19,  29,  73 },  // y = d207
+  {  77,   7,  64, 116, 132, 122,  37, 126, 120 },  // y = d63
+  { 101,  21, 107, 181, 192, 103,  19,  67, 125 }   // y = tm
+};
+
+const vp9_prob vp9_kf_partition_probs[PARTITION_CONTEXTS]
+                                     [PARTITION_TYPES - 1] = {
+  // 8x8 -> 4x4
+  { 158,  97,  94 },  // a/l both not split
+  {  93,  24,  99 },  // a split, l not split
+  {  85, 119,  44 },  // l split, a not split
+  {  62,  59,  67 },  // a/l both split
+  // 16x16 -> 8x8
+  { 149,  53,  53 },  // a/l both not split
+  {  94,  20,  48 },  // a split, l not split
+  {  83,  53,  24 },  // l split, a not split
+  {  52,  18,  18 },  // a/l both split
+  // 32x32 -> 16x16
+  { 150,  40,  39 },  // a/l both not split
+  {  78,  12,  26 },  // a split, l not split
+  {  67,  33,  11 },  // l split, a not split
+  {  24,   7,   5 },  // a/l both split
+  // 64x64 -> 32x32
+  { 174,  35,  49 },  // a/l both not split
+  {  68,  11,  27 },  // a split, l not split
+  {  57,  15,   9 },  // l split, a not split
+  {  12,   3,   3 },  // a/l both split
+};
+
+static const vp9_prob default_partition_probs[PARTITION_CONTEXTS]
+                                             [PARTITION_TYPES - 1] = {
+  // 8x8 -> 4x4
+  { 199, 122, 141 },  // a/l both not split
+  { 147,  63, 159 },  // a split, l not split
+  { 148, 133, 118 },  // l split, a not split
+  { 121, 104, 114 },  // a/l both split
+  // 16x16 -> 8x8
+  { 174,  73,  87 },  // a/l both not split
+  {  92,  41,  83 },  // a split, l not split
+  {  82,  99,  50 },  // l split, a not split
+  {  53,  39,  39 },  // a/l both split
+  // 32x32 -> 16x16
+  { 177,  58,  59 },  // a/l both not split
+  {  68,  26,  63 },  // a split, l not split
+  {  52,  79,  25 },  // l split, a not split
+  {  17,  14,  12 },  // a/l both split
+  // 64x64 -> 32x32
+  { 222,  34,  30 },  // a/l both not split
+  {  72,  16,  44 },  // a split, l not split
+  {  58,  32,  12 },  // l split, a not split
+  {  10,   7,   6 },  // a/l both split
+};
+
+static const vp9_prob default_inter_mode_probs[INTER_MODE_CONTEXTS]
+                                              [INTER_MODES - 1] = {
+  {2,       173,   34},  // 0 = both zero mv
+  {7,       145,   85},  // 1 = one zero mv + one a predicted mv
+  {7,       166,   63},  // 2 = two predicted mvs
+  {7,       94,    66},  // 3 = one predicted/zero and one new mv
+  {8,       64,    46},  // 4 = two new mvs
+  {17,      81,    31},  // 5 = one intra neighbour + x
+  {25,      29,    30},  // 6 = two intra neighbours
+};
+
+/* Array indices are identical to previously-existing INTRAMODECONTEXTNODES. */
+const vp9_tree_index vp9_intra_mode_tree[TREE_SIZE(INTRA_MODES)] = {
+  -DC_PRED, 2,                      /* 0 = DC_NODE */
+  -TM_PRED, 4,                      /* 1 = TM_NODE */
+  -V_PRED, 6,                       /* 2 = V_NODE */
+  8, 12,                            /* 3 = COM_NODE */
+  -H_PRED, 10,                      /* 4 = H_NODE */
+  -D135_PRED, -D117_PRED,           /* 5 = D135_NODE */
+  -D45_PRED, 14,                    /* 6 = D45_NODE */
+  -D63_PRED, 16,                    /* 7 = D63_NODE */
+  -D153_PRED, -D207_PRED             /* 8 = D153_NODE */
+};
+
+const vp9_tree_index vp9_inter_mode_tree[TREE_SIZE(INTER_MODES)] = {
+  -INTER_OFFSET(ZEROMV), 2,
+  -INTER_OFFSET(NEARESTMV), 4,
+  -INTER_OFFSET(NEARMV), -INTER_OFFSET(NEWMV)
+};
+
+const vp9_tree_index vp9_partition_tree[TREE_SIZE(PARTITION_TYPES)] = {
+  -PARTITION_NONE, 2,
+  -PARTITION_HORZ, 4,
+  -PARTITION_VERT, -PARTITION_SPLIT
+};
+
+static const vp9_prob default_intra_inter_p[INTRA_INTER_CONTEXTS] = {
+  9, 102, 187, 225
+};
+
+static const vp9_prob default_comp_inter_p[COMP_INTER_CONTEXTS] = {
+  239, 183, 119,  96,  41
+};
+
+static const vp9_prob default_comp_ref_p[REF_CONTEXTS] = {
+  50, 126, 123, 221, 226
+};
+
+static const vp9_prob default_single_ref_p[REF_CONTEXTS][2] = {
+  {  33,  16 },
+  {  77,  74 },
+  { 142, 142 },
+  { 172, 170 },
+  { 238, 247 }
+};
+
+static const struct tx_probs default_tx_probs = {
+  { { 3, 136, 37 },
+    { 5, 52,  13 } },
+
+  { { 20, 152 },
+    { 15, 101 } },
+
+  { { 100 },
+    { 66  } }
+};
+
+void tx_counts_to_branch_counts_32x32(const unsigned int *tx_count_32x32p,
+                                      unsigned int (*ct_32x32p)[2]) {
+  ct_32x32p[0][0] = tx_count_32x32p[TX_4X4];
+  ct_32x32p[0][1] = tx_count_32x32p[TX_8X8] +
+                    tx_count_32x32p[TX_16X16] +
+                    tx_count_32x32p[TX_32X32];
+  ct_32x32p[1][0] = tx_count_32x32p[TX_8X8];
+  ct_32x32p[1][1] = tx_count_32x32p[TX_16X16] +
+                    tx_count_32x32p[TX_32X32];
+  ct_32x32p[2][0] = tx_count_32x32p[TX_16X16];
+  ct_32x32p[2][1] = tx_count_32x32p[TX_32X32];
+}
+
+void tx_counts_to_branch_counts_16x16(const unsigned int *tx_count_16x16p,
+                                      unsigned int (*ct_16x16p)[2]) {
+  ct_16x16p[0][0] = tx_count_16x16p[TX_4X4];
+  ct_16x16p[0][1] = tx_count_16x16p[TX_8X8] + tx_count_16x16p[TX_16X16];
+  ct_16x16p[1][0] = tx_count_16x16p[TX_8X8];
+  ct_16x16p[1][1] = tx_count_16x16p[TX_16X16];
+}
+
+void tx_counts_to_branch_counts_8x8(const unsigned int *tx_count_8x8p,
+                                    unsigned int (*ct_8x8p)[2]) {
+  ct_8x8p[0][0] = tx_count_8x8p[TX_4X4];
+  ct_8x8p[0][1] = tx_count_8x8p[TX_8X8];
+}
+
+static const vp9_prob default_skip_probs[SKIP_CONTEXTS] = {
+  192, 128, 64
+};
+
+static const vp9_prob default_switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                                    [SWITCHABLE_FILTERS - 1] = {
+  { 235, 162, },
+  { 36, 255, },
+  { 34, 3, },
+  { 149, 144, },
+};
+
+void vp9_init_mode_probs(FRAME_CONTEXT *fc) {
+  vp9_copy(fc->uv_mode_prob, default_if_uv_probs);
+  vp9_copy(fc->y_mode_prob, default_if_y_probs);
+  vp9_copy(fc->switchable_interp_prob, default_switchable_interp_prob);
+  vp9_copy(fc->partition_prob, default_partition_probs);
+  vp9_copy(fc->intra_inter_prob, default_intra_inter_p);
+  vp9_copy(fc->comp_inter_prob, default_comp_inter_p);
+  vp9_copy(fc->comp_ref_prob, default_comp_ref_p);
+  vp9_copy(fc->single_ref_prob, default_single_ref_p);
+  fc->tx_probs = default_tx_probs;
+  vp9_copy(fc->skip_probs, default_skip_probs);
+  vp9_copy(fc->inter_mode_probs, default_inter_mode_probs);
+}
+
+const vp9_tree_index vp9_switchable_interp_tree
+                         [TREE_SIZE(SWITCHABLE_FILTERS)] = {
+  -EIGHTTAP, 2,
+  -EIGHTTAP_SMOOTH, -EIGHTTAP_SHARP
+};
+
+void vp9_adapt_mode_probs(VP9_COMMON *cm) {
+  int i, j;
+  FRAME_CONTEXT *fc = cm->fc;
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  const FRAME_COUNTS *counts = &cm->counts;
+
+  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+    fc->intra_inter_prob[i] = mode_mv_merge_probs(pre_fc->intra_inter_prob[i],
+                                                  counts->intra_inter[i]);
+  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+    fc->comp_inter_prob[i] = mode_mv_merge_probs(pre_fc->comp_inter_prob[i],
+                                                 counts->comp_inter[i]);
+  for (i = 0; i < REF_CONTEXTS; i++)
+    fc->comp_ref_prob[i] = mode_mv_merge_probs(pre_fc->comp_ref_prob[i],
+                                               counts->comp_ref[i]);
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      fc->single_ref_prob[i][j] = mode_mv_merge_probs(
+          pre_fc->single_ref_prob[i][j], counts->single_ref[i][j]);
+
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    vp9_tree_merge_probs(vp9_inter_mode_tree, pre_fc->inter_mode_probs[i],
+                counts->inter_mode[i], fc->inter_mode_probs[i]);
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
+    vp9_tree_merge_probs(vp9_intra_mode_tree, pre_fc->y_mode_prob[i],
+                counts->y_mode[i], fc->y_mode_prob[i]);
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    vp9_tree_merge_probs(vp9_intra_mode_tree, pre_fc->uv_mode_prob[i],
+                         counts->uv_mode[i], fc->uv_mode_prob[i]);
+
+  for (i = 0; i < PARTITION_CONTEXTS; i++)
+    vp9_tree_merge_probs(vp9_partition_tree, pre_fc->partition_prob[i],
+                         counts->partition[i], fc->partition_prob[i]);
+
+  if (cm->interp_filter == SWITCHABLE) {
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+      vp9_tree_merge_probs(vp9_switchable_interp_tree,
+                           pre_fc->switchable_interp_prob[i],
+                           counts->switchable_interp[i],
+                           fc->switchable_interp_prob[i]);
+  }
+
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    int j;
+    unsigned int branch_ct_8x8p[TX_SIZES - 3][2];
+    unsigned int branch_ct_16x16p[TX_SIZES - 2][2];
+    unsigned int branch_ct_32x32p[TX_SIZES - 1][2];
+
+    for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
+      tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], branch_ct_8x8p);
+      for (j = 0; j < TX_SIZES - 3; ++j)
+        fc->tx_probs.p8x8[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p8x8[i][j], branch_ct_8x8p[j]);
+
+      tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], branch_ct_16x16p);
+      for (j = 0; j < TX_SIZES - 2; ++j)
+        fc->tx_probs.p16x16[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p16x16[i][j], branch_ct_16x16p[j]);
+
+      tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], branch_ct_32x32p);
+      for (j = 0; j < TX_SIZES - 1; ++j)
+        fc->tx_probs.p32x32[i][j] = mode_mv_merge_probs(
+            pre_fc->tx_probs.p32x32[i][j], branch_ct_32x32p[j]);
+    }
+  }
+
+  for (i = 0; i < SKIP_CONTEXTS; ++i)
+    fc->skip_probs[i] = mode_mv_merge_probs(
+        pre_fc->skip_probs[i], counts->skip[i]);
+}
+
+static void set_default_lf_deltas(struct loopfilter *lf) {
+  lf->mode_ref_delta_enabled = 1;
+  lf->mode_ref_delta_update = 1;
+
+  lf->ref_deltas[INTRA_FRAME] = 1;
+  lf->ref_deltas[LAST_FRAME] = 0;
+  lf->ref_deltas[GOLDEN_FRAME] = -1;
+  lf->ref_deltas[ALTREF_FRAME] = -1;
+
+  lf->mode_deltas[0] = 0;
+  lf->mode_deltas[1] = 0;
+}
+
+void vp9_setup_past_independence(VP9_COMMON *cm) {
+  // Reset the segment feature data to the default stats:
+  // Features disabled, 0, with delta coding (Default state).
+  struct loopfilter *const lf = &cm->lf;
+
+  int i;
+  vp9_clearall_segfeatures(&cm->seg);
+  cm->seg.abs_delta = SEGMENT_DELTADATA;
+
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  if (cm->current_frame_seg_map)
+    memset(cm->current_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  // Reset the mode ref deltas for loop filter
+  vp9_zero(lf->last_ref_deltas);
+  vp9_zero(lf->last_mode_deltas);
+  set_default_lf_deltas(lf);
+
+  // To force update of the sharpness
+  lf->last_sharpness_level = -1;
+
+  vp9_default_coef_probs(cm);
+  vp9_init_mode_probs(cm->fc);
+  vp9_init_mv_probs(cm);
+  cm->fc->initialized = 1;
+
+  if (cm->frame_type == KEY_FRAME ||
+      cm->error_resilient_mode || cm->reset_frame_context == 3) {
+    // Reset all frame contexts.
+    for (i = 0; i < FRAME_CONTEXTS; ++i)
+      cm->frame_contexts[i] = *cm->fc;
+  } else if (cm->reset_frame_context == 2) {
+    // Reset only the frame context specified in the frame header.
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+  }
+
+  // prev_mip will only be allocated in encoder.
+  if (frame_is_intra_only(cm) && cm->prev_mip && !cm->frame_parallel_decode)
+    memset(cm->prev_mip, 0,
+           cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->prev_mip));
+
+  vp9_zero(cm->ref_frame_sign_bias);
+
+  cm->frame_context_idx = 0;
+}
diff --git a/media/libvpx/vp9/common/vp9_entropymode.h b/media/libvpx/vp9/common/vp9_entropymode.h
new file mode 100644
index 000000000..a0619ec6f
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropymode.h
@@ -0,0 +1,108 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENTROPYMODE_H_
+#define VP9_COMMON_VP9_ENTROPYMODE_H_
+
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BLOCK_SIZE_GROUPS 4
+
+#define TX_SIZE_CONTEXTS 2
+
+#define INTER_OFFSET(mode) ((mode) - NEARESTMV)
+
+struct VP9Common;
+
+struct tx_probs {
+  vp9_prob p32x32[TX_SIZE_CONTEXTS][TX_SIZES - 1];
+  vp9_prob p16x16[TX_SIZE_CONTEXTS][TX_SIZES - 2];
+  vp9_prob p8x8[TX_SIZE_CONTEXTS][TX_SIZES - 3];
+};
+
+struct tx_counts {
+  unsigned int p32x32[TX_SIZE_CONTEXTS][TX_SIZES];
+  unsigned int p16x16[TX_SIZE_CONTEXTS][TX_SIZES - 1];
+  unsigned int p8x8[TX_SIZE_CONTEXTS][TX_SIZES - 2];
+  unsigned int tx_totals[TX_SIZES];
+};
+
+typedef struct frame_contexts {
+  vp9_prob y_mode_prob[BLOCK_SIZE_GROUPS][INTRA_MODES - 1];
+  vp9_prob uv_mode_prob[INTRA_MODES][INTRA_MODES - 1];
+  vp9_prob partition_prob[PARTITION_CONTEXTS][PARTITION_TYPES - 1];
+  vp9_coeff_probs_model coef_probs[TX_SIZES][PLANE_TYPES];
+  vp9_prob switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                 [SWITCHABLE_FILTERS - 1];
+  vp9_prob inter_mode_probs[INTER_MODE_CONTEXTS][INTER_MODES - 1];
+  vp9_prob intra_inter_prob[INTRA_INTER_CONTEXTS];
+  vp9_prob comp_inter_prob[COMP_INTER_CONTEXTS];
+  vp9_prob single_ref_prob[REF_CONTEXTS][2];
+  vp9_prob comp_ref_prob[REF_CONTEXTS];
+  struct tx_probs tx_probs;
+  vp9_prob skip_probs[SKIP_CONTEXTS];
+  nmv_context nmvc;
+  int initialized;
+} FRAME_CONTEXT;
+
+typedef struct FRAME_COUNTS {
+  unsigned int y_mode[BLOCK_SIZE_GROUPS][INTRA_MODES];
+  unsigned int uv_mode[INTRA_MODES][INTRA_MODES];
+  unsigned int partition[PARTITION_CONTEXTS][PARTITION_TYPES];
+  vp9_coeff_count_model coef[TX_SIZES][PLANE_TYPES];
+  unsigned int eob_branch[TX_SIZES][PLANE_TYPES][REF_TYPES]
+                         [COEF_BANDS][COEFF_CONTEXTS];
+  unsigned int switchable_interp[SWITCHABLE_FILTER_CONTEXTS]
+                                [SWITCHABLE_FILTERS];
+  unsigned int inter_mode[INTER_MODE_CONTEXTS][INTER_MODES];
+  unsigned int intra_inter[INTRA_INTER_CONTEXTS][2];
+  unsigned int comp_inter[COMP_INTER_CONTEXTS][2];
+  unsigned int single_ref[REF_CONTEXTS][2][2];
+  unsigned int comp_ref[REF_CONTEXTS][2];
+  struct tx_counts tx;
+  unsigned int skip[SKIP_CONTEXTS][2];
+  nmv_context_counts mv;
+} FRAME_COUNTS;
+
+extern const vp9_prob vp9_kf_uv_mode_prob[INTRA_MODES][INTRA_MODES - 1];
+extern const vp9_prob vp9_kf_y_mode_prob[INTRA_MODES][INTRA_MODES]
+                                        [INTRA_MODES - 1];
+extern const vp9_prob vp9_kf_partition_probs[PARTITION_CONTEXTS]
+                                            [PARTITION_TYPES - 1];
+extern const vp9_tree_index vp9_intra_mode_tree[TREE_SIZE(INTRA_MODES)];
+extern const vp9_tree_index vp9_inter_mode_tree[TREE_SIZE(INTER_MODES)];
+extern const vp9_tree_index vp9_partition_tree[TREE_SIZE(PARTITION_TYPES)];
+extern const vp9_tree_index vp9_switchable_interp_tree
+                                [TREE_SIZE(SWITCHABLE_FILTERS)];
+
+void vp9_setup_past_independence(struct VP9Common *cm);
+
+void vp9_init_mode_probs(FRAME_CONTEXT *fc);
+
+void vp9_adapt_mode_probs(struct VP9Common *cm);
+
+void tx_counts_to_branch_counts_32x32(const unsigned int *tx_count_32x32p,
+                                      unsigned int (*ct_32x32p)[2]);
+void tx_counts_to_branch_counts_16x16(const unsigned int *tx_count_16x16p,
+                                      unsigned int (*ct_16x16p)[2]);
+void tx_counts_to_branch_counts_8x8(const unsigned int *tx_count_8x8p,
+                                    unsigned int (*ct_8x8p)[2]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPYMODE_H_
diff --git a/media/libvpx/vp9/common/vp9_entropymv.c b/media/libvpx/vp9/common/vp9_entropymv.c
new file mode 100644
index 000000000..2477e6ef3
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropymv.c
@@ -0,0 +1,222 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_entropymv.h"
+
+// Integer pel reference mv threshold for use of high-precision 1/8 mv
+#define COMPANDED_MVREF_THRESH 8
+
+const vp9_tree_index vp9_mv_joint_tree[TREE_SIZE(MV_JOINTS)] = {
+  -MV_JOINT_ZERO, 2,
+  -MV_JOINT_HNZVZ, 4,
+  -MV_JOINT_HZVNZ, -MV_JOINT_HNZVNZ
+};
+
+const vp9_tree_index vp9_mv_class_tree[TREE_SIZE(MV_CLASSES)] = {
+  -MV_CLASS_0, 2,
+  -MV_CLASS_1, 4,
+  6, 8,
+  -MV_CLASS_2, -MV_CLASS_3,
+  10, 12,
+  -MV_CLASS_4, -MV_CLASS_5,
+  -MV_CLASS_6, 14,
+  16, 18,
+  -MV_CLASS_7, -MV_CLASS_8,
+  -MV_CLASS_9, -MV_CLASS_10,
+};
+
+const vp9_tree_index vp9_mv_class0_tree[TREE_SIZE(CLASS0_SIZE)] = {
+  -0, -1,
+};
+
+const vp9_tree_index vp9_mv_fp_tree[TREE_SIZE(MV_FP_SIZE)] = {
+  -0, 2,
+  -1, 4,
+  -2, -3
+};
+
+static const nmv_context default_nmv_context = {
+  {32, 64, 96},
+  {
+    { // Vertical component
+      128,                                                  // sign
+      {224, 144, 192, 168, 192, 176, 192, 198, 198, 245},   // class
+      {216},                                                // class0
+      {136, 140, 148, 160, 176, 192, 224, 234, 234, 240},   // bits
+      {{128, 128, 64}, {96, 112, 64}},                      // class0_fp
+      {64, 96, 64},                                         // fp
+      160,                                                  // class0_hp bit
+      128,                                                  // hp
+    },
+    { // Horizontal component
+      128,                                                  // sign
+      {216, 128, 176, 160, 176, 176, 192, 198, 198, 208},   // class
+      {208},                                                // class0
+      {136, 140, 148, 160, 176, 192, 224, 234, 234, 240},   // bits
+      {{128, 128, 64}, {96, 112, 64}},                      // class0_fp
+      {64, 96, 64},                                         // fp
+      160,                                                  // class0_hp bit
+      128,                                                  // hp
+    }
+  },
+};
+
+static const uint8_t log_in_base_2[] = {
+  0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4,
+  4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10
+};
+
+static INLINE int mv_class_base(MV_CLASS_TYPE c) {
+  return c ? CLASS0_SIZE << (c + 2) : 0;
+}
+
+MV_CLASS_TYPE vp9_get_mv_class(int z, int *offset) {
+  const MV_CLASS_TYPE c = (z >= CLASS0_SIZE * 4096) ?
+      MV_CLASS_10 : (MV_CLASS_TYPE)log_in_base_2[z >> 3];
+  if (offset)
+    *offset = z - mv_class_base(c);
+  return c;
+}
+
+int vp9_use_mv_hp(const MV *ref) {
+  return (abs(ref->row) >> 3) < COMPANDED_MVREF_THRESH &&
+         (abs(ref->col) >> 3) < COMPANDED_MVREF_THRESH;
+}
+
+int vp9_get_mv_mag(MV_CLASS_TYPE c, int offset) {
+  return mv_class_base(c) + offset;
+}
+
+static void inc_mv_component(int v, nmv_component_counts *comp_counts,
+                             int incr, int usehp) {
+  int s, z, c, o, d, e, f;
+  assert(v != 0);            /* should not be zero */
+  s = v < 0;
+  comp_counts->sign[s] += incr;
+  z = (s ? -v : v) - 1;       /* magnitude - 1 */
+
+  c = vp9_get_mv_class(z, &o);
+  comp_counts->classes[c] += incr;
+
+  d = (o >> 3);               /* int mv data */
+  f = (o >> 1) & 3;           /* fractional pel mv data */
+  e = (o & 1);                /* high precision mv data */
+
+  if (c == MV_CLASS_0) {
+    comp_counts->class0[d] += incr;
+    comp_counts->class0_fp[d][f] += incr;
+    comp_counts->class0_hp[e] += usehp * incr;
+  } else {
+    int i;
+    int b = c + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < b; ++i)
+      comp_counts->bits[i][((d >> i) & 1)] += incr;
+    comp_counts->fp[f] += incr;
+    comp_counts->hp[e] += usehp * incr;
+  }
+}
+
+void vp9_inc_mv(const MV *mv, nmv_context_counts *counts) {
+  if (counts != NULL) {
+    const MV_JOINT_TYPE j = vp9_get_mv_joint(mv);
+    ++counts->joints[j];
+
+    if (mv_joint_vertical(j)) {
+      inc_mv_component(mv->row, &counts->comps[0], 1, 1);
+    }
+
+    if (mv_joint_horizontal(j)) {
+      inc_mv_component(mv->col, &counts->comps[1], 1, 1);
+    }
+  }
+}
+
+void vp9_adapt_mv_probs(VP9_COMMON *cm, int allow_hp) {
+  int i, j;
+
+  nmv_context *fc = &cm->fc->nmvc;
+  const nmv_context *pre_fc = &cm->frame_contexts[cm->frame_context_idx].nmvc;
+  const nmv_context_counts *counts = &cm->counts.mv;
+
+  vp9_tree_merge_probs(vp9_mv_joint_tree, pre_fc->joints, counts->joints,
+                       fc->joints);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *comp = &fc->comps[i];
+    const nmv_component *pre_comp = &pre_fc->comps[i];
+    const nmv_component_counts *c = &counts->comps[i];
+
+    comp->sign = mode_mv_merge_probs(pre_comp->sign, c->sign);
+    vp9_tree_merge_probs(vp9_mv_class_tree, pre_comp->classes, c->classes,
+                         comp->classes);
+    vp9_tree_merge_probs(vp9_mv_class0_tree, pre_comp->class0, c->class0,
+                         comp->class0);
+
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      comp->bits[j] = mode_mv_merge_probs(pre_comp->bits[j], c->bits[j]);
+
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      vp9_tree_merge_probs(vp9_mv_fp_tree, pre_comp->class0_fp[j],
+                           c->class0_fp[j], comp->class0_fp[j]);
+
+    vp9_tree_merge_probs(vp9_mv_fp_tree, pre_comp->fp, c->fp, comp->fp);
+
+    if (allow_hp) {
+      comp->class0_hp = mode_mv_merge_probs(pre_comp->class0_hp, c->class0_hp);
+      comp->hp = mode_mv_merge_probs(pre_comp->hp, c->hp);
+    }
+  }
+}
+
+void vp9_init_mv_probs(VP9_COMMON *cm) {
+  cm->fc->nmvc = default_nmv_context;
+}
diff --git a/media/libvpx/vp9/common/vp9_entropymv.h b/media/libvpx/vp9/common/vp9_entropymv.h
new file mode 100644
index 000000000..75e6861f4
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_entropymv.h
@@ -0,0 +1,134 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_ENTROPYMV_H_
+#define VP9_COMMON_VP9_ENTROPYMV_H_
+
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_mv.h"
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+
+void vp9_init_mv_probs(struct VP9Common *cm);
+
+void vp9_adapt_mv_probs(struct VP9Common *cm, int usehp);
+int vp9_use_mv_hp(const MV *ref);
+
+#define MV_UPDATE_PROB 252
+
+/* Symbols for coding which components are zero jointly */
+#define MV_JOINTS     4
+typedef enum {
+  MV_JOINT_ZERO = 0,             /* Zero vector */
+  MV_JOINT_HNZVZ = 1,            /* Vert zero, hor nonzero */
+  MV_JOINT_HZVNZ = 2,            /* Hor zero, vert nonzero */
+  MV_JOINT_HNZVNZ = 3,           /* Both components nonzero */
+} MV_JOINT_TYPE;
+
+static INLINE int mv_joint_vertical(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HZVNZ || type == MV_JOINT_HNZVNZ;
+}
+
+static INLINE int mv_joint_horizontal(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HNZVZ || type == MV_JOINT_HNZVNZ;
+}
+
+/* Symbols for coding magnitude class of nonzero components */
+#define MV_CLASSES     11
+typedef enum {
+  MV_CLASS_0 = 0,      /* (0, 2]     integer pel */
+  MV_CLASS_1 = 1,      /* (2, 4]     integer pel */
+  MV_CLASS_2 = 2,      /* (4, 8]     integer pel */
+  MV_CLASS_3 = 3,      /* (8, 16]    integer pel */
+  MV_CLASS_4 = 4,      /* (16, 32]   integer pel */
+  MV_CLASS_5 = 5,      /* (32, 64]   integer pel */
+  MV_CLASS_6 = 6,      /* (64, 128]  integer pel */
+  MV_CLASS_7 = 7,      /* (128, 256] integer pel */
+  MV_CLASS_8 = 8,      /* (256, 512] integer pel */
+  MV_CLASS_9 = 9,      /* (512, 1024] integer pel */
+  MV_CLASS_10 = 10,    /* (1024,2048] integer pel */
+} MV_CLASS_TYPE;
+
+#define CLASS0_BITS    1  /* bits at integer precision for class 0 */
+#define CLASS0_SIZE    (1 << CLASS0_BITS)
+#define MV_OFFSET_BITS (MV_CLASSES + CLASS0_BITS - 2)
+#define MV_FP_SIZE 4
+
+#define MV_MAX_BITS    (MV_CLASSES + CLASS0_BITS + 2)
+#define MV_MAX         ((1 << MV_MAX_BITS) - 1)
+#define MV_VALS        ((MV_MAX << 1) + 1)
+
+#define MV_IN_USE_BITS 14
+#define MV_UPP   ((1 << MV_IN_USE_BITS) - 1)
+#define MV_LOW   (-(1 << MV_IN_USE_BITS))
+
+extern const vp9_tree_index vp9_mv_joint_tree[];
+extern const vp9_tree_index vp9_mv_class_tree[];
+extern const vp9_tree_index vp9_mv_class0_tree[];
+extern const vp9_tree_index vp9_mv_fp_tree[];
+
+typedef struct {
+  vp9_prob sign;
+  vp9_prob classes[MV_CLASSES - 1];
+  vp9_prob class0[CLASS0_SIZE - 1];
+  vp9_prob bits[MV_OFFSET_BITS];
+  vp9_prob class0_fp[CLASS0_SIZE][MV_FP_SIZE - 1];
+  vp9_prob fp[MV_FP_SIZE - 1];
+  vp9_prob class0_hp;
+  vp9_prob hp;
+} nmv_component;
+
+typedef struct {
+  vp9_prob joints[MV_JOINTS - 1];
+  nmv_component comps[2];
+} nmv_context;
+
+static INLINE MV_JOINT_TYPE vp9_get_mv_joint(const MV *mv) {
+  if (mv->row == 0) {
+    return mv->col == 0 ? MV_JOINT_ZERO : MV_JOINT_HNZVZ;
+  } else {
+    return mv->col == 0 ? MV_JOINT_HZVNZ : MV_JOINT_HNZVNZ;
+  }
+}
+
+MV_CLASS_TYPE vp9_get_mv_class(int z, int *offset);
+int vp9_get_mv_mag(MV_CLASS_TYPE c, int offset);
+
+
+typedef struct {
+  unsigned int sign[2];
+  unsigned int classes[MV_CLASSES];
+  unsigned int class0[CLASS0_SIZE];
+  unsigned int bits[MV_OFFSET_BITS][2];
+  unsigned int class0_fp[CLASS0_SIZE][MV_FP_SIZE];
+  unsigned int fp[MV_FP_SIZE];
+  unsigned int class0_hp[2];
+  unsigned int hp[2];
+} nmv_component_counts;
+
+typedef struct {
+  unsigned int joints[MV_JOINTS];
+  nmv_component_counts comps[2];
+} nmv_context_counts;
+
+void vp9_inc_mv(const MV *mv, nmv_context_counts *mvctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENTROPYMV_H_
diff --git a/media/libvpx/vp9/common/vp9_enums.h b/media/libvpx/vp9/common/vp9_enums.h
new file mode 100644
index 000000000..048202593
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_enums.h
@@ -0,0 +1,149 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ENUMS_H_
+#define VP9_COMMON_VP9_ENUMS_H_
+
+#include "./vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MI_SIZE_LOG2 3
+#define MI_BLOCK_SIZE_LOG2 (6 - MI_SIZE_LOG2)  // 64 = 2^6
+
+#define MI_SIZE (1 << MI_SIZE_LOG2)  // pixels per mi-unit
+#define MI_BLOCK_SIZE (1 << MI_BLOCK_SIZE_LOG2)  // mi-units per max block
+
+#define MI_MASK (MI_BLOCK_SIZE - 1)
+
+// Bitstream profiles indicated by 2-3 bits in the uncompressed header.
+// 00: Profile 0.  8-bit 4:2:0 only.
+// 10: Profile 1.  8-bit 4:4:4, 4:2:2, and 4:4:0.
+// 01: Profile 2.  10-bit and 12-bit color only, with 4:2:0 sampling.
+// 110: Profile 3. 10-bit and 12-bit color only, with 4:2:2/4:4:4/4:4:0
+//                 sampling.
+// 111: Undefined profile.
+typedef enum BITSTREAM_PROFILE {
+  PROFILE_0,
+  PROFILE_1,
+  PROFILE_2,
+  PROFILE_3,
+  MAX_PROFILES
+} BITSTREAM_PROFILE;
+
+typedef enum BLOCK_SIZE {
+  BLOCK_4X4,
+  BLOCK_4X8,
+  BLOCK_8X4,
+  BLOCK_8X8,
+  BLOCK_8X16,
+  BLOCK_16X8,
+  BLOCK_16X16,
+  BLOCK_16X32,
+  BLOCK_32X16,
+  BLOCK_32X32,
+  BLOCK_32X64,
+  BLOCK_64X32,
+  BLOCK_64X64,
+  BLOCK_SIZES,
+  BLOCK_INVALID = BLOCK_SIZES
+} BLOCK_SIZE;
+
+typedef enum PARTITION_TYPE {
+  PARTITION_NONE,
+  PARTITION_HORZ,
+  PARTITION_VERT,
+  PARTITION_SPLIT,
+  PARTITION_TYPES,
+  PARTITION_INVALID = PARTITION_TYPES
+} PARTITION_TYPE;
+
+typedef char PARTITION_CONTEXT;
+#define PARTITION_PLOFFSET   4  // number of probability models per block size
+#define PARTITION_CONTEXTS (4 * PARTITION_PLOFFSET)
+
+// block transform size
+typedef enum {
+  TX_4X4 = 0,                      // 4x4 transform
+  TX_8X8 = 1,                      // 8x8 transform
+  TX_16X16 = 2,                    // 16x16 transform
+  TX_32X32 = 3,                    // 32x32 transform
+  TX_SIZES
+} TX_SIZE;
+
+// frame transform mode
+typedef enum {
+  ONLY_4X4            = 0,        // only 4x4 transform used
+  ALLOW_8X8           = 1,        // allow block transform size up to 8x8
+  ALLOW_16X16         = 2,        // allow block transform size up to 16x16
+  ALLOW_32X32         = 3,        // allow block transform size up to 32x32
+  TX_MODE_SELECT      = 4,        // transform specified for each block
+  TX_MODES            = 5,
+} TX_MODE;
+
+typedef enum {
+  DCT_DCT   = 0,                      // DCT  in both horizontal and vertical
+  ADST_DCT  = 1,                      // ADST in vertical, DCT in horizontal
+  DCT_ADST  = 2,                      // DCT  in vertical, ADST in horizontal
+  ADST_ADST = 3,                      // ADST in both directions
+  TX_TYPES = 4
+} TX_TYPE;
+
+typedef enum {
+  VP9_LAST_FLAG = 1 << 0,
+  VP9_GOLD_FLAG = 1 << 1,
+  VP9_ALT_FLAG = 1 << 2,
+} VP9_REFFRAME;
+
+typedef enum {
+  PLANE_TYPE_Y  = 0,
+  PLANE_TYPE_UV = 1,
+  PLANE_TYPES
+} PLANE_TYPE;
+
+typedef enum {
+  DC_PRED,         // Average of above and left pixels
+  V_PRED,          // Vertical
+  H_PRED,          // Horizontal
+  D45_PRED,        // Directional 45  deg = round(arctan(1/1) * 180/pi)
+  D135_PRED,       // Directional 135 deg = 180 - 45
+  D117_PRED,       // Directional 117 deg = 180 - 63
+  D153_PRED,       // Directional 153 deg = 180 - 27
+  D207_PRED,       // Directional 207 deg = 180 + 27
+  D63_PRED,        // Directional 63  deg = round(arctan(2/1) * 180/pi)
+  TM_PRED,         // True-motion
+  NEARESTMV,
+  NEARMV,
+  ZEROMV,
+  NEWMV,
+  MB_MODE_COUNT
+} PREDICTION_MODE;
+
+#define INTRA_MODES (TM_PRED + 1)
+
+#define INTER_MODES (1 + NEWMV - NEARESTMV)
+
+#define SKIP_CONTEXTS 3
+#define INTER_MODE_CONTEXTS 7
+
+/* Segment Feature Masks */
+#define MAX_MV_REF_CANDIDATES 2
+
+#define INTRA_INTER_CONTEXTS 4
+#define COMP_INTER_CONTEXTS 5
+#define REF_CONTEXTS 5
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ENUMS_H_
diff --git a/media/libvpx/vp9/common/vp9_filter.c b/media/libvpx/vp9/common/vp9_filter.c
new file mode 100644
index 000000000..b256d4af5
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_filter.c
@@ -0,0 +1,110 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_filter.h"
+
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_bilinear_filters[SUBPEL_SHIFTS]) = {
+  { 0, 0, 0, 128,   0, 0, 0, 0 },
+  { 0, 0, 0, 120,   8, 0, 0, 0 },
+  { 0, 0, 0, 112,  16, 0, 0, 0 },
+  { 0, 0, 0, 104,  24, 0, 0, 0 },
+  { 0, 0, 0,  96,  32, 0, 0, 0 },
+  { 0, 0, 0,  88,  40, 0, 0, 0 },
+  { 0, 0, 0,  80,  48, 0, 0, 0 },
+  { 0, 0, 0,  72,  56, 0, 0, 0 },
+  { 0, 0, 0,  64,  64, 0, 0, 0 },
+  { 0, 0, 0,  56,  72, 0, 0, 0 },
+  { 0, 0, 0,  48,  80, 0, 0, 0 },
+  { 0, 0, 0,  40,  88, 0, 0, 0 },
+  { 0, 0, 0,  32,  96, 0, 0, 0 },
+  { 0, 0, 0,  24, 104, 0, 0, 0 },
+  { 0, 0, 0,  16, 112, 0, 0, 0 },
+  { 0, 0, 0,   8, 120, 0, 0, 0 }
+};
+
+// Lagrangian interpolation filter
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8[SUBPEL_SHIFTS]) = {
+  { 0,   0,   0, 128,   0,   0,   0,  0},
+  { 0,   1,  -5, 126,   8,  -3,   1,  0},
+  { -1,   3, -10, 122,  18,  -6,   2,  0},
+  { -1,   4, -13, 118,  27,  -9,   3, -1},
+  { -1,   4, -16, 112,  37, -11,   4, -1},
+  { -1,   5, -18, 105,  48, -14,   4, -1},
+  { -1,   5, -19,  97,  58, -16,   5, -1},
+  { -1,   6, -19,  88,  68, -18,   5, -1},
+  { -1,   6, -19,  78,  78, -19,   6, -1},
+  { -1,   5, -18,  68,  88, -19,   6, -1},
+  { -1,   5, -16,  58,  97, -19,   5, -1},
+  { -1,   4, -14,  48, 105, -18,   5, -1},
+  { -1,   4, -11,  37, 112, -16,   4, -1},
+  { -1,   3,  -9,  27, 118, -13,   4, -1},
+  { 0,   2,  -6,  18, 122, -10,   3, -1},
+  { 0,   1,  -3,   8, 126,  -5,   1,  0}
+};
+
+// DCT based filter
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8s[SUBPEL_SHIFTS]) = {
+  {0,   0,   0, 128,   0,   0,   0, 0},
+  {-1,   3,  -7, 127,   8,  -3,   1, 0},
+  {-2,   5, -13, 125,  17,  -6,   3, -1},
+  {-3,   7, -17, 121,  27, -10,   5, -2},
+  {-4,   9, -20, 115,  37, -13,   6, -2},
+  {-4,  10, -23, 108,  48, -16,   8, -3},
+  {-4,  10, -24, 100,  59, -19,   9, -3},
+  {-4,  11, -24,  90,  70, -21,  10, -4},
+  {-4,  11, -23,  80,  80, -23,  11, -4},
+  {-4,  10, -21,  70,  90, -24,  11, -4},
+  {-3,   9, -19,  59, 100, -24,  10, -4},
+  {-3,   8, -16,  48, 108, -23,  10, -4},
+  {-2,   6, -13,  37, 115, -20,   9, -4},
+  {-2,   5, -10,  27, 121, -17,   7, -3},
+  {-1,   3,  -6,  17, 125, -13,   5, -2},
+  {0,   1,  -3,   8, 127,  -7,   3, -1}
+};
+
+// freqmultiplier = 0.5
+DECLARE_ALIGNED(256, const InterpKernel,
+                vp9_sub_pel_filters_8lp[SUBPEL_SHIFTS]) = {
+  { 0,  0,  0, 128,  0,  0,  0,  0},
+  {-3, -1, 32,  64, 38,  1, -3,  0},
+  {-2, -2, 29,  63, 41,  2, -3,  0},
+  {-2, -2, 26,  63, 43,  4, -4,  0},
+  {-2, -3, 24,  62, 46,  5, -4,  0},
+  {-2, -3, 21,  60, 49,  7, -4,  0},
+  {-1, -4, 18,  59, 51,  9, -4,  0},
+  {-1, -4, 16,  57, 53, 12, -4, -1},
+  {-1, -4, 14,  55, 55, 14, -4, -1},
+  {-1, -4, 12,  53, 57, 16, -4, -1},
+  { 0, -4,  9,  51, 59, 18, -4, -1},
+  { 0, -4,  7,  49, 60, 21, -3, -2},
+  { 0, -4,  5,  46, 62, 24, -3, -2},
+  { 0, -4,  4,  43, 63, 26, -2, -2},
+  { 0, -3,  2,  41, 63, 29, -2, -2},
+  { 0, -3,  1,  38, 64, 32, -1, -3}
+};
+
+
+static const InterpKernel* vp9_filter_kernels[4] = {
+  vp9_sub_pel_filters_8,
+  vp9_sub_pel_filters_8lp,
+  vp9_sub_pel_filters_8s,
+  vp9_bilinear_filters
+};
+
+const InterpKernel *vp9_get_interp_kernel(INTERP_FILTER filter) {
+  assert(filter != SWITCHABLE);
+  return vp9_filter_kernels[filter];
+}
+
diff --git a/media/libvpx/vp9/common/vp9_filter.h b/media/libvpx/vp9/common/vp9_filter.h
new file mode 100644
index 000000000..808a270fa
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_filter.h
@@ -0,0 +1,50 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_FILTER_H_
+#define VP9_COMMON_VP9_FILTER_H_
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+
+#define SUBPEL_BITS 4
+#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1)
+#define SUBPEL_SHIFTS (1 << SUBPEL_BITS)
+#define SUBPEL_TAPS 8
+
+typedef enum {
+  EIGHTTAP = 0,
+  EIGHTTAP_SMOOTH = 1,
+  EIGHTTAP_SHARP = 2,
+  SWITCHABLE_FILTERS = 3, /* Number of switchable filters */
+  BILINEAR = 3,
+  // The codec can operate in four possible inter prediction filter mode:
+  // 8-tap, 8-tap-smooth, 8-tap-sharp, and switching between the three.
+  SWITCHABLE_FILTER_CONTEXTS = SWITCHABLE_FILTERS + 1,
+  SWITCHABLE = 4  /* should be the last one */
+} INTERP_FILTER;
+
+typedef int16_t InterpKernel[SUBPEL_TAPS];
+
+const InterpKernel *vp9_get_interp_kernel(INTERP_FILTER filter);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_FILTER_H_
diff --git a/media/libvpx/vp9/common/vp9_frame_buffers.c b/media/libvpx/vp9/common/vp9_frame_buffers.c
new file mode 100644
index 000000000..0f41d6698
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_frame_buffers.c
@@ -0,0 +1,86 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_frame_buffers.h"
+#include "vpx_mem/vpx_mem.h"
+
+int vp9_alloc_internal_frame_buffers(InternalFrameBufferList *list) {
+  assert(list != NULL);
+  vp9_free_internal_frame_buffers(list);
+
+  list->num_internal_frame_buffers =
+      VP9_MAXIMUM_REF_BUFFERS + VPX_MAXIMUM_WORK_BUFFERS;
+  list->int_fb =
+      (InternalFrameBuffer *)vpx_calloc(list->num_internal_frame_buffers,
+                                        sizeof(*list->int_fb));
+  return (list->int_fb == NULL);
+}
+
+void vp9_free_internal_frame_buffers(InternalFrameBufferList *list) {
+  int i;
+
+  assert(list != NULL);
+
+  for (i = 0; i < list->num_internal_frame_buffers; ++i) {
+    vpx_free(list->int_fb[i].data);
+    list->int_fb[i].data = NULL;
+  }
+  vpx_free(list->int_fb);
+  list->int_fb = NULL;
+}
+
+int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
+                         vpx_codec_frame_buffer_t *fb) {
+  int i;
+  InternalFrameBufferList *const int_fb_list =
+      (InternalFrameBufferList *)cb_priv;
+  if (int_fb_list == NULL)
+    return -1;
+
+  // Find a free frame buffer.
+  for (i = 0; i < int_fb_list->num_internal_frame_buffers; ++i) {
+    if (!int_fb_list->int_fb[i].in_use)
+      break;
+  }
+
+  if (i == int_fb_list->num_internal_frame_buffers)
+    return -1;
+
+  if (int_fb_list->int_fb[i].size < min_size) {
+    int_fb_list->int_fb[i].data =
+        (uint8_t *)vpx_realloc(int_fb_list->int_fb[i].data, min_size);
+    if (!int_fb_list->int_fb[i].data)
+      return -1;
+
+    // This memset is needed for fixing valgrind error from C loop filter
+    // due to access uninitialized memory in frame border. It could be
+    // removed if border is totally removed.
+    memset(int_fb_list->int_fb[i].data, 0, min_size);
+    int_fb_list->int_fb[i].size = min_size;
+  }
+
+  fb->data = int_fb_list->int_fb[i].data;
+  fb->size = int_fb_list->int_fb[i].size;
+  int_fb_list->int_fb[i].in_use = 1;
+
+  // Set the frame buffer's private data to point at the internal frame buffer.
+  fb->priv = &int_fb_list->int_fb[i];
+  return 0;
+}
+
+int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb) {
+  InternalFrameBuffer *const int_fb = (InternalFrameBuffer *)fb->priv;
+  (void)cb_priv;
+  if (int_fb)
+    int_fb->in_use = 0;
+  return 0;
+}
diff --git a/media/libvpx/vp9/common/vp9_frame_buffers.h b/media/libvpx/vp9/common/vp9_frame_buffers.h
new file mode 100644
index 000000000..e2cfe61b6
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_frame_buffers.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_FRAME_BUFFERS_H_
+#define VP9_COMMON_VP9_FRAME_BUFFERS_H_
+
+#include "vpx/vpx_frame_buffer.h"
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct InternalFrameBuffer {
+  uint8_t *data;
+  size_t size;
+  int in_use;
+} InternalFrameBuffer;
+
+typedef struct InternalFrameBufferList {
+  int num_internal_frame_buffers;
+  InternalFrameBuffer *int_fb;
+} InternalFrameBufferList;
+
+// Initializes |list|. Returns 0 on success.
+int vp9_alloc_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Free any data allocated to the frame buffers.
+void vp9_free_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Callback used by libvpx to request an external frame buffer. |cb_priv|
+// Callback private data, which points to an InternalFrameBufferList.
+// |min_size| is the minimum size in bytes needed to decode the next frame.
+// |fb| pointer to the frame buffer.
+int vp9_get_frame_buffer(void *cb_priv, size_t min_size,
+                         vpx_codec_frame_buffer_t *fb);
+
+// Callback used by libvpx when there are no references to the frame buffer.
+// |cb_priv| is not used. |fb| pointer to the frame buffer.
+int vp9_release_frame_buffer(void *cb_priv, vpx_codec_frame_buffer_t *fb);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_FRAME_BUFFERS_H_
diff --git a/media/libvpx/vp9/common/vp9_idct.c b/media/libvpx/vp9/common/vp9_idct.c
new file mode 100644
index 000000000..174b96e21
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_idct.c
@@ -0,0 +1,2867 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_systemdependent.h"
+
+static INLINE uint8_t clip_pixel_add(uint8_t dest, tran_high_t trans) {
+  trans = WRAPLOW(trans, 8);
+  return clip_pixel(WRAPLOW(dest + trans, 8));
+}
+
+void vp9_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+/* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+   0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1, 8);
+    op[1] = WRAPLOW(b1, 8);
+    op[2] = WRAPLOW(c1, 8);
+    op[3] = WRAPLOW(d1, 8);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = clip_pixel_add(dest[stride * 0], a1);
+    dest[stride * 1] = clip_pixel_add(dest[stride * 1], b1);
+    dest[stride * 2] = clip_pixel_add(dest[stride * 2], c1);
+    dest[stride * 3] = clip_pixel_add(dest[stride * 3], d1);
+
+    ip++;
+    dest++;
+  }
+}
+
+void vp9_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest, int dest_stride) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1, 8);
+  op[1] = op[2] = op[3] = WRAPLOW(e1, 8);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = clip_pixel_add(dest[dest_stride * 0], a1);
+    dest[dest_stride * 1] = clip_pixel_add(dest[dest_stride * 1], e1);
+    dest[dest_stride * 2] = clip_pixel_add(dest[dest_stride * 2], e1);
+    dest[dest_stride * 3] = clip_pixel_add(dest[dest_stride * 3], e1);
+    ip++;
+    dest++;
+  }
+}
+
+static void idct4(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3], 8);
+  output[1] = WRAPLOW(step[1] + step[2], 8);
+  output[2] = WRAPLOW(step[1] - step[2], 8);
+  output[3] = WRAPLOW(step[0] - step[3], 8);
+}
+
+void vp9_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    idct4(input, outptr);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    idct4(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+void vp9_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest,
+                         int dest_stride) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = clip_pixel_add(dest[0], a1);
+    dest[1] = clip_pixel_add(dest[1], a1);
+    dest[2] = clip_pixel_add(dest[2], a1);
+    dest[3] = clip_pixel_add(dest[3], a1);
+    dest += dest_stride;
+  }
+}
+
+static void idct8(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2 & stage 3 - even half
+  idct4(step1, step1);
+
+  // stage 2 - odd half
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  // stage 3 -odd half
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7], 8);
+  output[1] = WRAPLOW(step1[1] + step1[6], 8);
+  output[2] = WRAPLOW(step1[2] + step1[5], 8);
+  output[3] = WRAPLOW(step1[3] + step1[4], 8);
+  output[4] = WRAPLOW(step1[3] - step1[4], 8);
+  output[5] = WRAPLOW(step1[2] - step1[5], 8);
+  output[6] = WRAPLOW(step1[1] - step1[6], 8);
+  output[7] = WRAPLOW(step1[0] - step1[7], 8);
+}
+
+void vp9_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  for (i = 0; i < 8; ++i) {
+    idct8(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    idct8(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void vp9_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+static void iadst4(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = x0 - x2 + x3;
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(dct_const_round_shift(s0 + s3), 8);
+  output[1] = WRAPLOW(dct_const_round_shift(s1 + s3), 8);
+  output[2] = WRAPLOW(dct_const_round_shift(s2), 8);
+  output[3] = WRAPLOW(dct_const_round_shift(s0 + s1 - s3), 8);
+}
+
+void vp9_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  const transform_2d IHT_4[] = {
+    { idct4, idct4  },  // DCT_DCT  = 0
+    { iadst4, idct4  },   // ADST_DCT = 1
+    { idct4, iadst4 },    // DCT_ADST = 2
+    { iadst4, iadst4 }      // ADST_ADST = 3
+  };
+
+  int i, j;
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // inverse transform row vectors
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].rows(input, outptr);
+    input  += 4;
+    outptr += 4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    IHT_4[tx_type].cols(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+static void iadst8(const tran_low_t *input, tran_low_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = (int)(cospi_2_64  * x0 + cospi_30_64 * x1);
+  s1 = (int)(cospi_30_64 * x0 - cospi_2_64  * x1);
+  s2 = (int)(cospi_10_64 * x2 + cospi_22_64 * x3);
+  s3 = (int)(cospi_22_64 * x2 - cospi_10_64 * x3);
+  s4 = (int)(cospi_18_64 * x4 + cospi_14_64 * x5);
+  s5 = (int)(cospi_14_64 * x4 - cospi_18_64 * x5);
+  s6 = (int)(cospi_26_64 * x6 + cospi_6_64  * x7);
+  s7 = (int)(cospi_6_64  * x6 - cospi_26_64 * x7);
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s4), 8);
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s5), 8);
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s6), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s7), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s0 - s4), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s1 - s5), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s2 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s3 - s7), 8);
+
+  // stage 2
+  s0 = (int)x0;
+  s1 = (int)x1;
+  s2 = (int)x2;
+  s3 = (int)x3;
+  s4 = (int)(cospi_8_64 * x4 + cospi_24_64 * x5);
+  s5 = (int)(cospi_24_64 * x4 - cospi_8_64 * x5);
+  s6 = (int)(-cospi_24_64 * x6 + cospi_8_64 * x7);
+  s7 = (int)(cospi_8_64 * x6 + cospi_24_64 * x7);
+
+  x0 = WRAPLOW(s0 + s2, 8);
+  x1 = WRAPLOW(s1 + s3, 8);
+  x2 = WRAPLOW(s0 - s2, 8);
+  x3 = WRAPLOW(s1 - s3, 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7), 8);
+
+  // stage 3
+  s2 = (int)(cospi_16_64 * (x2 + x3));
+  s3 = (int)(cospi_16_64 * (x2 - x3));
+  s6 = (int)(cospi_16_64 * (x6 + x7));
+  s7 = (int)(cospi_16_64 * (x6 - x7));
+
+  x2 = WRAPLOW(dct_const_round_shift(s2), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7), 8);
+
+  output[0] = WRAPLOW(x0, 8);
+  output[1] = WRAPLOW(-x4, 8);
+  output[2] = WRAPLOW(x6, 8);
+  output[3] = WRAPLOW(-x2, 8);
+  output[4] = WRAPLOW(x3, 8);
+  output[5] = WRAPLOW(-x7, 8);
+  output[6] = WRAPLOW(x5, 8);
+  output[7] = WRAPLOW(-x1, 8);
+}
+
+static const transform_2d IHT_8[] = {
+  { idct8,  idct8  },  // DCT_DCT  = 0
+  { iadst8, idct8  },  // ADST_DCT = 1
+  { idct8,  iadst8 },  // DCT_ADST = 2
+  { iadst8, iadst8 }   // ADST_ADST = 3
+};
+
+void vp9_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  int i, j;
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[8], temp_out[8];
+  const transform_2d ht = IHT_8[tx_type];
+
+  // inverse transform row vectors
+  for (i = 0; i < 8; ++i) {
+    ht.rows(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void vp9_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  // only first 4 row has non-zero coefs
+  for (i = 0; i < 4; ++i) {
+    idct8(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    idct8(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+static void idct16(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], 8);
+  step1[9] = WRAPLOW(step2[8] - step2[9], 8);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], 8);
+  step1[11] = WRAPLOW(step2[10] + step2[11], 8);
+  step1[12] = WRAPLOW(step2[12] + step2[13], 8);
+  step1[13] = WRAPLOW(step2[12] - step2[13], 8);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], 8);
+  step1[15] = WRAPLOW(step2[14] + step2[15], 8);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[2], 8);
+  step1[2] = WRAPLOW(step2[1] - step2[2], 8);
+  step1[3] = WRAPLOW(step2[0] - step2[3], 8);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], 8);
+  step1[9] = WRAPLOW(step2[9] + step2[10], 8);
+  step1[10] = WRAPLOW(step2[9] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[8] - step2[11], 8);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], 8);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], 8);
+  step1[14] = WRAPLOW(step2[13] + step2[14], 8);
+  step1[15] = WRAPLOW(step2[12] + step2[15], 8);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], 8);
+  step2[1] = WRAPLOW(step1[1] + step1[6], 8);
+  step2[2] = WRAPLOW(step1[2] + step1[5], 8);
+  step2[3] = WRAPLOW(step1[3] + step1[4], 8);
+  step2[4] = WRAPLOW(step1[3] - step1[4], 8);
+  step2[5] = WRAPLOW(step1[2] - step1[5], 8);
+  step2[6] = WRAPLOW(step1[1] - step1[6], 8);
+  step2[7] = WRAPLOW(step1[0] - step1[7], 8);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15], 8);
+  output[1] = WRAPLOW(step2[1] + step2[14], 8);
+  output[2] = WRAPLOW(step2[2] + step2[13], 8);
+  output[3] = WRAPLOW(step2[3] + step2[12], 8);
+  output[4] = WRAPLOW(step2[4] + step2[11], 8);
+  output[5] = WRAPLOW(step2[5] + step2[10], 8);
+  output[6] = WRAPLOW(step2[6] + step2[9], 8);
+  output[7] = WRAPLOW(step2[7] + step2[8], 8);
+  output[8] = WRAPLOW(step2[7] - step2[8], 8);
+  output[9] = WRAPLOW(step2[6] - step2[9], 8);
+  output[10] = WRAPLOW(step2[5] - step2[10], 8);
+  output[11] = WRAPLOW(step2[4] - step2[11], 8);
+  output[12] = WRAPLOW(step2[3] - step2[12], 8);
+  output[13] = WRAPLOW(step2[2] - step2[13], 8);
+  output[14] = WRAPLOW(step2[1] - step2[14], 8);
+  output[15] = WRAPLOW(step2[0] - step2[15], 8);
+}
+
+void vp9_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows
+  for (i = 0; i < 16; ++i) {
+    idct16(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    idct16(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+static void iadst16(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4]
+              = output[5] = output[6] = output[7] = output[8]
+              = output[9] = output[10] = output[11] = output[12]
+              = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s8), 8);
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s9), 8);
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s10), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s11), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s12), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s13), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6 + s14), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7 + s15), 8);
+  x8 = WRAPLOW(dct_const_round_shift(s0 - s8), 8);
+  x9 = WRAPLOW(dct_const_round_shift(s1 - s9), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s2 - s10), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s3 - s11), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s4 - s12), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s5 - s13), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s6 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s7 - s15), 8);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 =    x8 * cospi_4_64   + x9 * cospi_28_64;
+  s9 =    x8 * cospi_28_64  - x9 * cospi_4_64;
+  s10 =   x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 =   x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = - x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 =   x12 * cospi_4_64  + x13 * cospi_28_64;
+  s14 = - x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 =   x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4, 8);
+  x1 = WRAPLOW(s1 + s5, 8);
+  x2 = WRAPLOW(s2 + s6, 8);
+  x3 = WRAPLOW(s3 + s7, 8);
+  x4 = WRAPLOW(s0 - s4, 8);
+  x5 = WRAPLOW(s1 - s5, 8);
+  x6 = WRAPLOW(s2 - s6, 8);
+  x7 = WRAPLOW(s3 - s7, 8);
+  x8 = WRAPLOW(dct_const_round_shift(s8 + s12), 8);
+  x9 = WRAPLOW(dct_const_round_shift(s9 + s13), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s10 + s14), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s11 + s15), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s8 - s12), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s9 - s13), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s10 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s11 - s15), 8);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64  + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = - x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 =   x6 * cospi_8_64  + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64  + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = - x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 =   x14 * cospi_8_64  + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(check_range(s0 + s2), 8);
+  x1 = WRAPLOW(check_range(s1 + s3), 8);
+  x2 = WRAPLOW(check_range(s0 - s2), 8);
+  x3 = WRAPLOW(check_range(s1 - s3), 8);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6), 8);
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7), 8);
+  x8 = WRAPLOW(check_range(s8 + s10), 8);
+  x9 = WRAPLOW(check_range(s9 + s11), 8);
+  x10 = WRAPLOW(check_range(s8 - s10), 8);
+  x11 = WRAPLOW(check_range(s9 - s11), 8);
+  x12 = WRAPLOW(dct_const_round_shift(s12 + s14), 8);
+  x13 = WRAPLOW(dct_const_round_shift(s13 + s15), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s12 - s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s13 - s15), 8);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (- x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (- x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(dct_const_round_shift(s2), 8);
+  x3 = WRAPLOW(dct_const_round_shift(s3), 8);
+  x6 = WRAPLOW(dct_const_round_shift(s6), 8);
+  x7 = WRAPLOW(dct_const_round_shift(s7), 8);
+  x10 = WRAPLOW(dct_const_round_shift(s10), 8);
+  x11 = WRAPLOW(dct_const_round_shift(s11), 8);
+  x14 = WRAPLOW(dct_const_round_shift(s14), 8);
+  x15 = WRAPLOW(dct_const_round_shift(s15), 8);
+
+  output[0] = WRAPLOW(x0, 8);
+  output[1] = WRAPLOW(-x8, 8);
+  output[2] = WRAPLOW(x12, 8);
+  output[3] = WRAPLOW(-x4, 8);
+  output[4] = WRAPLOW(x6, 8);
+  output[5] = WRAPLOW(x14, 8);
+  output[6] = WRAPLOW(x10, 8);
+  output[7] = WRAPLOW(x2, 8);
+  output[8] = WRAPLOW(x3, 8);
+  output[9] = WRAPLOW(x11, 8);
+  output[10] = WRAPLOW(x15, 8);
+  output[11] = WRAPLOW(x7, 8);
+  output[12] = WRAPLOW(x5, 8);
+  output[13] = WRAPLOW(-x13, 8);
+  output[14] = WRAPLOW(x9, 8);
+  output[15] = WRAPLOW(-x1, 8);
+}
+
+static const transform_2d IHT_16[] = {
+  { idct16,  idct16  },  // DCT_DCT  = 0
+  { iadst16, idct16  },  // ADST_DCT = 1
+  { idct16,  iadst16 },  // DCT_ADST = 2
+  { iadst16, iadst16 }   // ADST_ADST = 3
+};
+
+void vp9_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  int i, j;
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[16], temp_out[16];
+  const transform_2d ht = IHT_16[tx_type];
+
+  // Rows
+  for (i = 0; i < 16; ++i) {
+    ht.rows(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp9_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    idct16(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    idct16(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp9_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+static void idct32(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[31] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step2[16] = WRAPLOW(step1[16] + step1[17], 8);
+  step2[17] = WRAPLOW(step1[16] - step1[17], 8);
+  step2[18] = WRAPLOW(-step1[18] + step1[19], 8);
+  step2[19] = WRAPLOW(step1[18] + step1[19], 8);
+  step2[20] = WRAPLOW(step1[20] + step1[21], 8);
+  step2[21] = WRAPLOW(step1[20] - step1[21], 8);
+  step2[22] = WRAPLOW(-step1[22] + step1[23], 8);
+  step2[23] = WRAPLOW(step1[22] + step1[23], 8);
+  step2[24] = WRAPLOW(step1[24] + step1[25], 8);
+  step2[25] = WRAPLOW(step1[24] - step1[25], 8);
+  step2[26] = WRAPLOW(-step1[26] + step1[27], 8);
+  step2[27] = WRAPLOW(step1[26] + step1[27], 8);
+  step2[28] = WRAPLOW(step1[28] + step1[29], 8);
+  step2[29] = WRAPLOW(step1[28] - step1[29], 8);
+  step2[30] = WRAPLOW(-step1[30] + step1[31], 8);
+  step2[31] = WRAPLOW(step1[30] + step1[31], 8);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], 8);
+  step1[9] = WRAPLOW(step2[8] - step2[9], 8);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], 8);
+  step1[11] = WRAPLOW(step2[10] + step2[11], 8);
+  step1[12] = WRAPLOW(step2[12] + step2[13], 8);
+  step1[13] = WRAPLOW(step2[12] - step2[13], 8);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], 8);
+  step1[15] = WRAPLOW(step2[14] + step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[4] = WRAPLOW(step1[4] + step1[5], 8);
+  step2[5] = WRAPLOW(step1[4] - step1[5], 8);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], 8);
+  step2[7] = WRAPLOW(step1[6] + step1[7], 8);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19], 8);
+  step2[17] = WRAPLOW(step1[17] + step1[18], 8);
+  step2[18] = WRAPLOW(step1[17] - step1[18], 8);
+  step2[19] = WRAPLOW(step1[16] - step1[19], 8);
+  step2[20] = WRAPLOW(-step1[20] + step1[23], 8);
+  step2[21] = WRAPLOW(-step1[21] + step1[22], 8);
+  step2[22] = WRAPLOW(step1[21] + step1[22], 8);
+  step2[23] = WRAPLOW(step1[20] + step1[23], 8);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27], 8);
+  step2[25] = WRAPLOW(step1[25] + step1[26], 8);
+  step2[26] = WRAPLOW(step1[25] - step1[26], 8);
+  step2[27] = WRAPLOW(step1[24] - step1[27], 8);
+  step2[28] = WRAPLOW(-step1[28] + step1[31], 8);
+  step2[29] = WRAPLOW(-step1[29] + step1[30], 8);
+  step2[30] = WRAPLOW(step1[29] + step1[30], 8);
+  step2[31] = WRAPLOW(step1[28] + step1[31], 8);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[2], 8);
+  step1[2] = WRAPLOW(step2[1] - step2[2], 8);
+  step1[3] = WRAPLOW(step2[0] - step2[3], 8);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], 8);
+  step1[9] = WRAPLOW(step2[9] + step2[10], 8);
+  step1[10] = WRAPLOW(step2[9] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[8] - step2[11], 8);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], 8);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], 8);
+  step1[14] = WRAPLOW(step2[13] + step2[14], 8);
+  step1[15] = WRAPLOW(step2[12] + step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], 8);
+  step2[1] = WRAPLOW(step1[1] + step1[6], 8);
+  step2[2] = WRAPLOW(step1[2] + step1[5], 8);
+  step2[3] = WRAPLOW(step1[3] + step1[4], 8);
+  step2[4] = WRAPLOW(step1[3] - step1[4], 8);
+  step2[5] = WRAPLOW(step1[2] - step1[5], 8);
+  step2[6] = WRAPLOW(step1[1] - step1[6], 8);
+  step2[7] = WRAPLOW(step1[0] - step1[7], 8);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23], 8);
+  step2[17] = WRAPLOW(step1[17] + step1[22], 8);
+  step2[18] = WRAPLOW(step1[18] + step1[21], 8);
+  step2[19] = WRAPLOW(step1[19] + step1[20], 8);
+  step2[20] = WRAPLOW(step1[19] - step1[20], 8);
+  step2[21] = WRAPLOW(step1[18] - step1[21], 8);
+  step2[22] = WRAPLOW(step1[17] - step1[22], 8);
+  step2[23] = WRAPLOW(step1[16] - step1[23], 8);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31], 8);
+  step2[25] = WRAPLOW(-step1[25] + step1[30], 8);
+  step2[26] = WRAPLOW(-step1[26] + step1[29], 8);
+  step2[27] = WRAPLOW(-step1[27] + step1[28], 8);
+  step2[28] = WRAPLOW(step1[27] + step1[28], 8);
+  step2[29] = WRAPLOW(step1[26] + step1[29], 8);
+  step2[30] = WRAPLOW(step1[25] + step1[30], 8);
+  step2[31] = WRAPLOW(step1[24] + step1[31], 8);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15], 8);
+  step1[1] = WRAPLOW(step2[1] + step2[14], 8);
+  step1[2] = WRAPLOW(step2[2] + step2[13], 8);
+  step1[3] = WRAPLOW(step2[3] + step2[12], 8);
+  step1[4] = WRAPLOW(step2[4] + step2[11], 8);
+  step1[5] = WRAPLOW(step2[5] + step2[10], 8);
+  step1[6] = WRAPLOW(step2[6] + step2[9], 8);
+  step1[7] = WRAPLOW(step2[7] + step2[8], 8);
+  step1[8] = WRAPLOW(step2[7] - step2[8], 8);
+  step1[9] = WRAPLOW(step2[6] - step2[9], 8);
+  step1[10] = WRAPLOW(step2[5] - step2[10], 8);
+  step1[11] = WRAPLOW(step2[4] - step2[11], 8);
+  step1[12] = WRAPLOW(step2[3] - step2[12], 8);
+  step1[13] = WRAPLOW(step2[2] - step2[13], 8);
+  step1[14] = WRAPLOW(step2[1] - step2[14], 8);
+  step1[15] = WRAPLOW(step2[0] - step2[15], 8);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1), 8);
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2), 8);
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31], 8);
+  output[1] = WRAPLOW(step1[1] + step1[30], 8);
+  output[2] = WRAPLOW(step1[2] + step1[29], 8);
+  output[3] = WRAPLOW(step1[3] + step1[28], 8);
+  output[4] = WRAPLOW(step1[4] + step1[27], 8);
+  output[5] = WRAPLOW(step1[5] + step1[26], 8);
+  output[6] = WRAPLOW(step1[6] + step1[25], 8);
+  output[7] = WRAPLOW(step1[7] + step1[24], 8);
+  output[8] = WRAPLOW(step1[8] + step1[23], 8);
+  output[9] = WRAPLOW(step1[9] + step1[22], 8);
+  output[10] = WRAPLOW(step1[10] + step1[21], 8);
+  output[11] = WRAPLOW(step1[11] + step1[20], 8);
+  output[12] = WRAPLOW(step1[12] + step1[19], 8);
+  output[13] = WRAPLOW(step1[13] + step1[18], 8);
+  output[14] = WRAPLOW(step1[14] + step1[17], 8);
+  output[15] = WRAPLOW(step1[15] + step1[16], 8);
+  output[16] = WRAPLOW(step1[15] - step1[16], 8);
+  output[17] = WRAPLOW(step1[14] - step1[17], 8);
+  output[18] = WRAPLOW(step1[13] - step1[18], 8);
+  output[19] = WRAPLOW(step1[12] - step1[19], 8);
+  output[20] = WRAPLOW(step1[11] - step1[20], 8);
+  output[21] = WRAPLOW(step1[10] - step1[21], 8);
+  output[22] = WRAPLOW(step1[9] - step1[22], 8);
+  output[23] = WRAPLOW(step1[8] - step1[23], 8);
+  output[24] = WRAPLOW(step1[7] - step1[24], 8);
+  output[25] = WRAPLOW(step1[6] - step1[25], 8);
+  output[26] = WRAPLOW(step1[5] - step1[26], 8);
+  output[27] = WRAPLOW(step1[4] - step1[27], 8);
+  output[28] = WRAPLOW(step1[3] - step1[28], 8);
+  output[29] = WRAPLOW(step1[2] - step1[29], 8);
+  output[30] = WRAPLOW(step1[1] - step1[30], 8);
+  output[31] = WRAPLOW(step1[0] - step1[31], 8);
+}
+
+void vp9_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    int16_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      idct32(input, outptr);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    idct32(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp9_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 8x8 has non-zero coeff
+  for (i = 0; i < 8; ++i) {
+    idct32(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    idct32(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void vp9_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), 8);
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64), 8);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+// idct
+void vp9_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    vp9_idct4x4_16_add(input, dest, stride);
+  else
+    vp9_idct4x4_1_add(input, dest, stride);
+}
+
+
+void vp9_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    vp9_iwht4x4_16_add(input, dest, stride);
+  else
+    vp9_iwht4x4_1_add(input, dest, stride);
+}
+
+void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  // If dc is 1, then input[0] is the reconstructed value, do not need
+  // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1.
+
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to decide what to do.
+  // TODO(yunqingwang): "eobs = 1" case is also handled in vp9_short_idct8x8_c.
+  // Combine that with code here.
+  if (eob == 1)
+    // DC only DCT coefficient
+    vp9_idct8x8_1_add(input, dest, stride);
+  else if (eob <= 12)
+    vp9_idct8x8_12_add(input, dest, stride);
+  else
+    vp9_idct8x8_64_add(input, dest, stride);
+}
+
+void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob) {
+  /* The calculation can be simplified if there are not many non-zero dct
+   * coefficients. Use eobs to separate different cases. */
+  if (eob == 1)
+    /* DC only DCT coefficient. */
+    vp9_idct16x16_1_add(input, dest, stride);
+  else if (eob <= 10)
+    vp9_idct16x16_10_add(input, dest, stride);
+  else
+    vp9_idct16x16_256_add(input, dest, stride);
+}
+
+void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob) {
+  if (eob == 1)
+    vp9_idct32x32_1_add(input, dest, stride);
+  else if (eob <= 34)
+    // non-zero coeff only in upper-left 8x8
+    vp9_idct32x32_34_add(input, dest, stride);
+  else
+    vp9_idct32x32_1024_add(input, dest, stride);
+}
+
+// iht
+void vp9_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob) {
+  if (tx_type == DCT_DCT)
+    vp9_idct4x4_add(input, dest, stride, eob);
+  else
+    vp9_iht4x4_16_add(input, dest, stride, tx_type);
+}
+
+void vp9_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob) {
+  if (tx_type == DCT_DCT) {
+    vp9_idct8x8_add(input, dest, stride, eob);
+  } else {
+    vp9_iht8x8_64_add(input, dest, stride, tx_type);
+  }
+}
+
+void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                      int stride, int eob) {
+  if (tx_type == DCT_DCT) {
+    vp9_idct16x16_add(input, dest, stride, eob);
+  } else {
+    vp9_iht16x16_256_add(input, dest, stride, tx_type);
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  /* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+     0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1, bd);
+    op[1] = WRAPLOW(b1, bd);
+    op[2] = WRAPLOW(c1, bd);
+    op[3] = WRAPLOW(d1, bd);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = highbd_clip_pixel_add(dest[stride * 0], a1, bd);
+    dest[stride * 1] = highbd_clip_pixel_add(dest[stride * 1], b1, bd);
+    dest[stride * 2] = highbd_clip_pixel_add(dest[stride * 2], c1, bd);
+    dest[stride * 3] = highbd_clip_pixel_add(dest[stride * 3], d1, bd);
+
+    ip++;
+    dest++;
+  }
+}
+
+void vp9_highbd_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  (void) bd;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1, bd);
+  op[1] = op[2] = op[3] = WRAPLOW(e1, bd);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = highbd_clip_pixel_add(
+        dest[dest_stride * 0], a1, bd);
+    dest[dest_stride * 1] = highbd_clip_pixel_add(
+        dest[dest_stride * 1], e1, bd);
+    dest[dest_stride * 2] = highbd_clip_pixel_add(
+        dest[dest_stride * 2], e1, bd);
+    dest[dest_stride * 3] = highbd_clip_pixel_add(
+        dest[dest_stride * 3], e1, bd);
+    ip++;
+    dest++;
+  }
+}
+
+void vp9_highbd_idct4(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  (void) bd;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3], bd);
+  output[1] = WRAPLOW(step[1] + step[2], bd);
+  output[2] = WRAPLOW(step[1] - step[2], bd);
+  output[3] = WRAPLOW(step[0] - step[3], bd);
+}
+
+void vp9_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    vp9_highbd_idct4(input, outptr, bd);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    vp9_highbd_idct4(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = highbd_clip_pixel_add(dest[0], a1, bd);
+    dest[1] = highbd_clip_pixel_add(dest[1], a1, bd);
+    dest[2] = highbd_clip_pixel_add(dest[2], a1, bd);
+    dest[3] = highbd_clip_pixel_add(dest[3], a1, bd);
+    dest += dest_stride;
+  }
+}
+
+void vp9_highbd_idct8(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2 & stage 3 - even half
+  vp9_highbd_idct4(step1, step1, bd);
+
+  // stage 2 - odd half
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  // stage 3 - odd half
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7], bd);
+  output[1] = WRAPLOW(step1[1] + step1[6], bd);
+  output[2] = WRAPLOW(step1[2] + step1[5], bd);
+  output[3] = WRAPLOW(step1[3] + step1[4], bd);
+  output[4] = WRAPLOW(step1[3] - step1[4], bd);
+  output[5] = WRAPLOW(step1[2] - step1[5], bd);
+  output[6] = WRAPLOW(step1[1] - step1[6], bd);
+  output[7] = WRAPLOW(step1[0] - step1[7], bd);
+}
+
+void vp9_highbd_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 8; ++i) {
+    vp9_highbd_idct8(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp9_highbd_idct8(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+static void highbd_iadst4(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3)) {
+    memset(output, 0, 4 * sizeof(*output));
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = (tran_high_t)(x0 - x2 + x3);
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(highbd_dct_const_round_shift(s0 + s3, bd), bd);
+  output[1] = WRAPLOW(highbd_dct_const_round_shift(s1 + s3, bd), bd);
+  output[2] = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  output[3] = WRAPLOW(highbd_dct_const_round_shift(s0 + s1 - s3, bd), bd);
+}
+
+void vp9_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  const highbd_transform_2d IHT_4[] = {
+    { vp9_highbd_idct4, vp9_highbd_idct4  },    // DCT_DCT  = 0
+    { highbd_iadst4, vp9_highbd_idct4 },    // ADST_DCT = 1
+    { vp9_highbd_idct4, highbd_iadst4 },    // DCT_ADST = 2
+    { highbd_iadst4, highbd_iadst4 }    // ADST_ADST = 3
+  };
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Inverse transform row vectors.
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].rows(input, outptr, bd);
+    input  += 4;
+    outptr += 4;
+  }
+
+  // Inverse transform column vectors.
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j)
+      temp_in[j] = out[j * 4 + i];
+    IHT_4[tx_type].cols(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+static void highbd_iadst8(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[7];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[5];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[3];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[1];
+  tran_low_t x7 = input[6];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    memset(output, 0, 8 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = cospi_2_64  * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64  * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64  * x7;
+  s7 = cospi_6_64  * x6 - cospi_26_64 * x7;
+
+  x0 = WRAPLOW(highbd_dct_const_round_shift(s0 + s4, bd), bd);
+  x1 = WRAPLOW(highbd_dct_const_round_shift(s1 + s5, bd), bd);
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2 + s6, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3 + s7, bd), bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s0 - s4, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s1 - s5, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s2 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s3 - s7, bd), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 =  cospi_8_64  * x4 + cospi_24_64 * x5;
+  s5 =  cospi_24_64 * x4 - cospi_8_64  * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64  * x7;
+  s7 =  cospi_8_64  * x6 + cospi_24_64 * x7;
+
+  x0 = WRAPLOW(s0 + s2, bd);
+  x1 = WRAPLOW(s1 + s3, bd);
+  x2 = WRAPLOW(s0 - s2, bd);
+  x3 = WRAPLOW(s1 - s3, bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s6, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s7, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s4 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s5 - s7, bd), bd);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7, bd), bd);
+
+  output[0] = WRAPLOW(x0, bd);
+  output[1] = WRAPLOW(-x4, bd);
+  output[2] = WRAPLOW(x6, bd);
+  output[3] = WRAPLOW(-x2, bd);
+  output[4] = WRAPLOW(x3, bd);
+  output[5] = WRAPLOW(-x7, bd);
+  output[6] = WRAPLOW(x5, bd);
+  output[7] = WRAPLOW(-x1, bd);
+}
+
+static const highbd_transform_2d HIGH_IHT_8[] = {
+  { vp9_highbd_idct8,  vp9_highbd_idct8  },  // DCT_DCT  = 0
+  { highbd_iadst8, vp9_highbd_idct8  },  // ADST_DCT = 1
+  { vp9_highbd_idct8,  highbd_iadst8 },  // DCT_ADST = 2
+  { highbd_iadst8, highbd_iadst8 }   // ADST_ADST = 3
+};
+
+void vp9_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  int i, j;
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[8], temp_out[8];
+  const highbd_transform_2d ht = HIGH_IHT_8[tx_type];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Inverse transform row vectors.
+  for (i = 0; i < 8; ++i) {
+    ht.rows(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Inverse transform column vectors.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    ht.cols(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct8x8_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  // Only first 4 row has non-zero coefs.
+  for (i = 0; i < 4; ++i) {
+    vp9_highbd_idct8(input, outptr, bd);
+    input += 8;
+    outptr += 8;
+  }
+  // Then transform columns.
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j)
+      temp_in[j] = out[j * 8 + i];
+    vp9_highbd_idct8(temp_in, temp_out, bd);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct16(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0/2];
+  step1[1] = input[16/2];
+  step1[2] = input[8/2];
+  step1[3] = input[24/2];
+  step1[4] = input[4/2];
+  step1[5] = input[20/2];
+  step1[6] = input[12/2];
+  step1[7] = input[28/2];
+  step1[8] = input[2/2];
+  step1[9] = input[18/2];
+  step1[10] = input[10/2];
+  step1[11] = input[26/2];
+  step1[12] = input[6/2];
+  step1[13] = input[22/2];
+  step1[14] = input[14/2];
+  step1[15] = input[30/2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[15] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = WRAPLOW(step2[14] + step2[15], bd);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = WRAPLOW(step2[12] + step2[15], bd);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15], bd);
+  output[1] = WRAPLOW(step2[1] + step2[14], bd);
+  output[2] = WRAPLOW(step2[2] + step2[13], bd);
+  output[3] = WRAPLOW(step2[3] + step2[12], bd);
+  output[4] = WRAPLOW(step2[4] + step2[11], bd);
+  output[5] = WRAPLOW(step2[5] + step2[10], bd);
+  output[6] = WRAPLOW(step2[6] + step2[9], bd);
+  output[7] = WRAPLOW(step2[7] + step2[8], bd);
+  output[8] = WRAPLOW(step2[7] - step2[8], bd);
+  output[9] = WRAPLOW(step2[6] - step2[9], bd);
+  output[10] = WRAPLOW(step2[5] - step2[10], bd);
+  output[11] = WRAPLOW(step2[4] - step2[11], bd);
+  output[12] = WRAPLOW(step2[3] - step2[12], bd);
+  output[13] = WRAPLOW(step2[2] - step2[13], bd);
+  output[14] = WRAPLOW(step2[1] - step2[14], bd);
+  output[15] = WRAPLOW(step2[0] - step2[15], bd);
+}
+
+void vp9_highbd_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows.
+  for (i = 0; i < 16; ++i) {
+    vp9_highbd_idct16(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    vp9_highbd_idct16(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+static void highbd_iadst16(const tran_low_t *input, tran_low_t *output,
+                           int bd) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_low_t x0 = input[15];
+  tran_low_t x1 = input[0];
+  tran_low_t x2 = input[13];
+  tran_low_t x3 = input[2];
+  tran_low_t x4 = input[11];
+  tran_low_t x5 = input[4];
+  tran_low_t x6 = input[9];
+  tran_low_t x7 = input[6];
+  tran_low_t x8 = input[7];
+  tran_low_t x9 = input[8];
+  tran_low_t x10 = input[5];
+  tran_low_t x11 = input[10];
+  tran_low_t x12 = input[3];
+  tran_low_t x13 = input[12];
+  tran_low_t x14 = input[1];
+  tran_low_t x15 = input[14];
+  (void) bd;
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8
+           | x9 | x10 | x11 | x12 | x13 | x14 | x15)) {
+    memset(output, 0, 16 * sizeof(*output));
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64  + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64  + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64  + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64  - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64  - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(highbd_dct_const_round_shift(s0 + s8, bd), bd);
+  x1 = WRAPLOW(highbd_dct_const_round_shift(s1 + s9, bd), bd);
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2 + s10, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3 + s11, bd), bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s12, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s13, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6 + s14, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7 + s15, bd), bd);
+  x8  = WRAPLOW(highbd_dct_const_round_shift(s0 - s8, bd), bd);
+  x9  = WRAPLOW(highbd_dct_const_round_shift(s1 - s9, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s2 - s10, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s3 - s11, bd), bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s4 - s12, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s5 - s13, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s6 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s7 - s15, bd), bd);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4, bd);
+  x1 = WRAPLOW(s1 + s5, bd);
+  x2 = WRAPLOW(s2 + s6, bd);
+  x3 = WRAPLOW(s3 + s7, bd);
+  x4 = WRAPLOW(s0 - s4, bd);
+  x5 = WRAPLOW(s1 - s5, bd);
+  x6 = WRAPLOW(s2 - s6, bd);
+  x7 = WRAPLOW(s3 - s7, bd);
+  x8 = WRAPLOW(highbd_dct_const_round_shift(s8 + s12, bd), bd);
+  x9 = WRAPLOW(highbd_dct_const_round_shift(s9 + s13, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s10 + s14, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s11 + s15, bd), bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s8 - s12, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s9 - s13, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s10 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s11 - s15, bd), bd);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(s0 + s2, bd);
+  x1 = WRAPLOW(s1 + s3, bd);
+  x2 = WRAPLOW(s0 - s2, bd);
+  x3 = WRAPLOW(s1 - s3, bd);
+  x4 = WRAPLOW(highbd_dct_const_round_shift(s4 + s6, bd), bd);
+  x5 = WRAPLOW(highbd_dct_const_round_shift(s5 + s7, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s4 - s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s5 - s7, bd), bd);
+  x8 = WRAPLOW(s8 + s10, bd);
+  x9 = WRAPLOW(s9 + s11, bd);
+  x10 = WRAPLOW(s8 - s10, bd);
+  x11 = WRAPLOW(s9 - s11, bd);
+  x12 = WRAPLOW(highbd_dct_const_round_shift(s12 + s14, bd), bd);
+  x13 = WRAPLOW(highbd_dct_const_round_shift(s13 + s15, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s12 - s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s13 - s15, bd), bd);
+
+  // stage 4
+  s2 = (- cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (- cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(highbd_dct_const_round_shift(s2, bd), bd);
+  x3 = WRAPLOW(highbd_dct_const_round_shift(s3, bd), bd);
+  x6 = WRAPLOW(highbd_dct_const_round_shift(s6, bd), bd);
+  x7 = WRAPLOW(highbd_dct_const_round_shift(s7, bd), bd);
+  x10 = WRAPLOW(highbd_dct_const_round_shift(s10, bd), bd);
+  x11 = WRAPLOW(highbd_dct_const_round_shift(s11, bd), bd);
+  x14 = WRAPLOW(highbd_dct_const_round_shift(s14, bd), bd);
+  x15 = WRAPLOW(highbd_dct_const_round_shift(s15, bd), bd);
+
+  output[0] = WRAPLOW(x0, bd);
+  output[1] = WRAPLOW(-x8, bd);
+  output[2] = WRAPLOW(x12, bd);
+  output[3] = WRAPLOW(-x4, bd);
+  output[4] = WRAPLOW(x6, bd);
+  output[5] = WRAPLOW(x14, bd);
+  output[6] = WRAPLOW(x10, bd);
+  output[7] = WRAPLOW(x2, bd);
+  output[8] = WRAPLOW(x3, bd);
+  output[9] = WRAPLOW(x11, bd);
+  output[10] = WRAPLOW(x15, bd);
+  output[11] = WRAPLOW(x7, bd);
+  output[12] = WRAPLOW(x5, bd);
+  output[13] = WRAPLOW(-x13, bd);
+  output[14] = WRAPLOW(x9, bd);
+  output[15] = WRAPLOW(-x1, bd);
+}
+
+static const highbd_transform_2d HIGH_IHT_16[] = {
+  { vp9_highbd_idct16,  vp9_highbd_idct16  },  // DCT_DCT  = 0
+  { highbd_iadst16, vp9_highbd_idct16  },  // ADST_DCT = 1
+  { vp9_highbd_idct16,  highbd_iadst16 },  // DCT_ADST = 2
+  { highbd_iadst16, highbd_iadst16 }   // ADST_ADST = 3
+};
+
+void vp9_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int tx_type, int bd) {
+  int i, j;
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[16], temp_out[16];
+  const highbd_transform_2d ht = HIGH_IHT_16[tx_type];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 16; ++i) {
+    ht.rows(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j * 16 + i];
+    ht.cols(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    vp9_highbd_idct16(input, outptr, bd);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns.
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j)
+      temp_in[j] = out[j*16 + i];
+    vp9_highbd_idct16(temp_in, temp_out, bd);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+static void highbd_idct32(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+  (void) bd;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[31] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[30] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[28] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[24] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[15] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step2[16] = WRAPLOW(step1[16] + step1[17], bd);
+  step2[17] = WRAPLOW(step1[16] - step1[17], bd);
+  step2[18] = WRAPLOW(-step1[18] + step1[19], bd);
+  step2[19] = WRAPLOW(step1[18] + step1[19], bd);
+  step2[20] = WRAPLOW(step1[20] + step1[21], bd);
+  step2[21] = WRAPLOW(step1[20] - step1[21], bd);
+  step2[22] = WRAPLOW(-step1[22] + step1[23], bd);
+  step2[23] = WRAPLOW(step1[22] + step1[23], bd);
+  step2[24] = WRAPLOW(step1[24] + step1[25], bd);
+  step2[25] = WRAPLOW(step1[24] - step1[25], bd);
+  step2[26] = WRAPLOW(-step1[26] + step1[27], bd);
+  step2[27] = WRAPLOW(step1[26] + step1[27], bd);
+  step2[28] = WRAPLOW(step1[28] + step1[29], bd);
+  step2[29] = WRAPLOW(step1[28] - step1[29], bd);
+  step2[30] = WRAPLOW(-step1[30] + step1[31], bd);
+  step2[31] = WRAPLOW(step1[30] + step1[31], bd);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[7] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+
+  step1[8] = WRAPLOW(step2[8] + step2[9], bd);
+  step1[9] = WRAPLOW(step2[8] - step2[9], bd);
+  step1[10] = WRAPLOW(-step2[10] + step2[11], bd);
+  step1[11] = WRAPLOW(step2[10] + step2[11], bd);
+  step1[12] = WRAPLOW(step2[12] + step2[13], bd);
+  step1[13] = WRAPLOW(step2[12] - step2[13], bd);
+  step1[14] = WRAPLOW(-step2[14] + step2[15], bd);
+  step1[15] = WRAPLOW(step2[14] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[30] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[1] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[3] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[4] = WRAPLOW(step1[4] + step1[5], bd);
+  step2[5] = WRAPLOW(step1[4] - step1[5], bd);
+  step2[6] = WRAPLOW(-step1[6] + step1[7], bd);
+  step2[7] = WRAPLOW(step1[6] + step1[7], bd);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[14] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19], bd);
+  step2[17] = WRAPLOW(step1[17] + step1[18], bd);
+  step2[18] = WRAPLOW(step1[17] - step1[18], bd);
+  step2[19] = WRAPLOW(step1[16] - step1[19], bd);
+  step2[20] = WRAPLOW(-step1[20] + step1[23], bd);
+  step2[21] = WRAPLOW(-step1[21] + step1[22], bd);
+  step2[22] = WRAPLOW(step1[21] + step1[22], bd);
+  step2[23] = WRAPLOW(step1[20] + step1[23], bd);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27], bd);
+  step2[25] = WRAPLOW(step1[25] + step1[26], bd);
+  step2[26] = WRAPLOW(step1[25] - step1[26], bd);
+  step2[27] = WRAPLOW(step1[24] - step1[27], bd);
+  step2[28] = WRAPLOW(-step1[28] + step1[31], bd);
+  step2[29] = WRAPLOW(-step1[29] + step1[30], bd);
+  step2[30] = WRAPLOW(step1[29] + step1[30], bd);
+  step2[31] = WRAPLOW(step1[28] + step1[31], bd);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[2], bd);
+  step1[2] = WRAPLOW(step2[1] - step2[2], bd);
+  step1[3] = WRAPLOW(step2[0] - step2[3], bd);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[6] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11], bd);
+  step1[9] = WRAPLOW(step2[9] + step2[10], bd);
+  step1[10] = WRAPLOW(step2[9] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[8] - step2[11], bd);
+  step1[12] = WRAPLOW(-step2[12] + step2[15], bd);
+  step1[13] = WRAPLOW(-step2[13] + step2[14], bd);
+  step1[14] = WRAPLOW(step2[13] + step2[14], bd);
+  step1[15] = WRAPLOW(step2[12] + step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[29] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[28] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7], bd);
+  step2[1] = WRAPLOW(step1[1] + step1[6], bd);
+  step2[2] = WRAPLOW(step1[2] + step1[5], bd);
+  step2[3] = WRAPLOW(step1[3] + step1[4], bd);
+  step2[4] = WRAPLOW(step1[3] - step1[4], bd);
+  step2[5] = WRAPLOW(step1[2] - step1[5], bd);
+  step2[6] = WRAPLOW(step1[1] - step1[6], bd);
+  step2[7] = WRAPLOW(step1[0] - step1[7], bd);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[13] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step2[12] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23], bd);
+  step2[17] = WRAPLOW(step1[17] + step1[22], bd);
+  step2[18] = WRAPLOW(step1[18] + step1[21], bd);
+  step2[19] = WRAPLOW(step1[19] + step1[20], bd);
+  step2[20] = WRAPLOW(step1[19] - step1[20], bd);
+  step2[21] = WRAPLOW(step1[18] - step1[21], bd);
+  step2[22] = WRAPLOW(step1[17] - step1[22], bd);
+  step2[23] = WRAPLOW(step1[16] - step1[23], bd);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31], bd);
+  step2[25] = WRAPLOW(-step1[25] + step1[30], bd);
+  step2[26] = WRAPLOW(-step1[26] + step1[29], bd);
+  step2[27] = WRAPLOW(-step1[27] + step1[28], bd);
+  step2[28] = WRAPLOW(step1[27] + step1[28], bd);
+  step2[29] = WRAPLOW(step1[26] + step1[29], bd);
+  step2[30] = WRAPLOW(step1[25] + step1[30], bd);
+  step2[31] = WRAPLOW(step1[24] + step1[31], bd);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15], bd);
+  step1[1] = WRAPLOW(step2[1] + step2[14], bd);
+  step1[2] = WRAPLOW(step2[2] + step2[13], bd);
+  step1[3] = WRAPLOW(step2[3] + step2[12], bd);
+  step1[4] = WRAPLOW(step2[4] + step2[11], bd);
+  step1[5] = WRAPLOW(step2[5] + step2[10], bd);
+  step1[6] = WRAPLOW(step2[6] + step2[9], bd);
+  step1[7] = WRAPLOW(step2[7] + step2[8], bd);
+  step1[8] = WRAPLOW(step2[7] - step2[8], bd);
+  step1[9] = WRAPLOW(step2[6] - step2[9], bd);
+  step1[10] = WRAPLOW(step2[5] - step2[10], bd);
+  step1[11] = WRAPLOW(step2[4] - step2[11], bd);
+  step1[12] = WRAPLOW(step2[3] - step2[12], bd);
+  step1[13] = WRAPLOW(step2[2] - step2[13], bd);
+  step1[14] = WRAPLOW(step2[1] - step2[14], bd);
+  step1[15] = WRAPLOW(step2[0] - step2[15], bd);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[27] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[26] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[25] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(highbd_dct_const_round_shift(temp1, bd), bd);
+  step1[24] = WRAPLOW(highbd_dct_const_round_shift(temp2, bd), bd);
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31], bd);
+  output[1] = WRAPLOW(step1[1] + step1[30], bd);
+  output[2] = WRAPLOW(step1[2] + step1[29], bd);
+  output[3] = WRAPLOW(step1[3] + step1[28], bd);
+  output[4] = WRAPLOW(step1[4] + step1[27], bd);
+  output[5] = WRAPLOW(step1[5] + step1[26], bd);
+  output[6] = WRAPLOW(step1[6] + step1[25], bd);
+  output[7] = WRAPLOW(step1[7] + step1[24], bd);
+  output[8] = WRAPLOW(step1[8] + step1[23], bd);
+  output[9] = WRAPLOW(step1[9] + step1[22], bd);
+  output[10] = WRAPLOW(step1[10] + step1[21], bd);
+  output[11] = WRAPLOW(step1[11] + step1[20], bd);
+  output[12] = WRAPLOW(step1[12] + step1[19], bd);
+  output[13] = WRAPLOW(step1[13] + step1[18], bd);
+  output[14] = WRAPLOW(step1[14] + step1[17], bd);
+  output[15] = WRAPLOW(step1[15] + step1[16], bd);
+  output[16] = WRAPLOW(step1[15] - step1[16], bd);
+  output[17] = WRAPLOW(step1[14] - step1[17], bd);
+  output[18] = WRAPLOW(step1[13] - step1[18], bd);
+  output[19] = WRAPLOW(step1[12] - step1[19], bd);
+  output[20] = WRAPLOW(step1[11] - step1[20], bd);
+  output[21] = WRAPLOW(step1[10] - step1[21], bd);
+  output[22] = WRAPLOW(step1[9] - step1[22], bd);
+  output[23] = WRAPLOW(step1[8] - step1[23], bd);
+  output[24] = WRAPLOW(step1[7] - step1[24], bd);
+  output[25] = WRAPLOW(step1[6] - step1[25], bd);
+  output[26] = WRAPLOW(step1[5] - step1[26], bd);
+  output[27] = WRAPLOW(step1[4] - step1[27], bd);
+  output[28] = WRAPLOW(step1[3] - step1[28], bd);
+  output[29] = WRAPLOW(step1[2] - step1[29], bd);
+  output[30] = WRAPLOW(step1[1] - step1[30], bd);
+  output[31] = WRAPLOW(step1[0] - step1[31], bd);
+}
+
+void vp9_highbd_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest8,
+                                     int stride, int bd) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_low_t zero_coeff[16];
+    for (j = 0; j < 16; ++j)
+      zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      highbd_idct32(input, outptr, bd);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int bd) {
+  tran_low_t out[32 * 32] = {0};
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  // Only upper-left 8x8 has non-zero coeff.
+  for (i = 0; i < 8; ++i) {
+    highbd_idct32(input, outptr, bd);
+    input += 32;
+    outptr += 32;
+  }
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j)
+      temp_in[j] = out[j * 32 + i];
+    highbd_idct32(temp_in, temp_out, bd);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+    }
+  }
+}
+
+void vp9_highbd_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bd) {
+  int i, j;
+  int a1;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  tran_low_t out = WRAPLOW(
+      highbd_dct_const_round_shift(input[0] * cospi_16_64, bd), bd);
+  out = WRAPLOW(highbd_dct_const_round_shift(out * cospi_16_64, bd), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i)
+      dest[i] = highbd_clip_pixel_add(dest[i], a1, bd);
+    dest += stride;
+  }
+}
+
+// idct
+void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    vp9_highbd_idct4x4_16_add(input, dest, stride, bd);
+  else
+    vp9_highbd_idct4x4_1_add(input, dest, stride, bd);
+}
+
+
+void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    vp9_highbd_iwht4x4_16_add(input, dest, stride, bd);
+  else
+    vp9_highbd_iwht4x4_1_add(input, dest, stride, bd);
+}
+
+void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  // If dc is 1, then input[0] is the reconstructed value, do not need
+  // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1.
+
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to decide what to do.
+  // TODO(yunqingwang): "eobs = 1" case is also handled in vp9_short_idct8x8_c.
+  // Combine that with code here.
+  // DC only DCT coefficient
+  if (eob == 1) {
+    vp9_highbd_idct8x8_1_add(input, dest, stride, bd);
+  } else if (eob <= 10) {
+    vp9_highbd_idct8x8_10_add(input, dest, stride, bd);
+  } else {
+    vp9_highbd_idct8x8_64_add(input, dest, stride, bd);
+  }
+}
+
+void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd) {
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to separate different cases.
+  // DC only DCT coefficient.
+  if (eob == 1) {
+    vp9_highbd_idct16x16_1_add(input, dest, stride, bd);
+  } else if (eob <= 10) {
+    vp9_highbd_idct16x16_10_add(input, dest, stride, bd);
+  } else {
+    vp9_highbd_idct16x16_256_add(input, dest, stride, bd);
+  }
+}
+
+void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd) {
+  // Non-zero coeff only in upper-left 8x8
+  if (eob == 1) {
+    vp9_highbd_idct32x32_1_add(input, dest, stride, bd);
+  } else if (eob <= 34) {
+    vp9_highbd_idct32x32_34_add(input, dest, stride, bd);
+  } else {
+    vp9_highbd_idct32x32_1024_add(input, dest, stride, bd);
+  }
+}
+
+// iht
+void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT)
+    vp9_highbd_idct4x4_add(input, dest, stride, eob, bd);
+  else
+    vp9_highbd_iht4x4_16_add(input, dest, stride, tx_type, bd);
+}
+
+void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_idct8x8_add(input, dest, stride, eob, bd);
+  } else {
+    vp9_highbd_iht8x8_64_add(input, dest, stride, tx_type, bd);
+  }
+}
+
+void vp9_highbd_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_idct16x16_add(input, dest, stride, eob, bd);
+  } else {
+    vp9_highbd_iht16x16_256_add(input, dest, stride, tx_type, bd);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/media/libvpx/vp9/common/vp9_idct.h b/media/libvpx/vp9/common/vp9_idct.h
new file mode 100644
index 000000000..cee1682a6
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_idct.h
@@ -0,0 +1,212 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_IDCT_H_
+#define VP9_COMMON_VP9_IDCT_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Constants and Macros used by all idct/dct functions
+#define DCT_CONST_BITS 14
+#define DCT_CONST_ROUNDING  (1 << (DCT_CONST_BITS - 1))
+
+#define UNIT_QUANT_SHIFT 2
+#define UNIT_QUANT_FACTOR (1 << UNIT_QUANT_SHIFT)
+
+#define pair_set_epi16(a, b) \
+  _mm_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define dual_set_epi16(a, b) \
+  _mm_set_epi16((int16_t)(b), (int16_t)(b), (int16_t)(b), (int16_t)(b), \
+                (int16_t)(a), (int16_t)(a), (int16_t)(a), (int16_t)(a))
+
+// Constants:
+//  for (int i = 1; i< 32; ++i)
+//    printf("static const int cospi_%d_64 = %.0f;\n", i,
+//           round(16384 * cos(i*M_PI/64)));
+// Note: sin(k*Pi/64) = cos((32-k)*Pi/64)
+static const tran_high_t cospi_1_64  = 16364;
+static const tran_high_t cospi_2_64  = 16305;
+static const tran_high_t cospi_3_64  = 16207;
+static const tran_high_t cospi_4_64  = 16069;
+static const tran_high_t cospi_5_64  = 15893;
+static const tran_high_t cospi_6_64  = 15679;
+static const tran_high_t cospi_7_64  = 15426;
+static const tran_high_t cospi_8_64  = 15137;
+static const tran_high_t cospi_9_64  = 14811;
+static const tran_high_t cospi_10_64 = 14449;
+static const tran_high_t cospi_11_64 = 14053;
+static const tran_high_t cospi_12_64 = 13623;
+static const tran_high_t cospi_13_64 = 13160;
+static const tran_high_t cospi_14_64 = 12665;
+static const tran_high_t cospi_15_64 = 12140;
+static const tran_high_t cospi_16_64 = 11585;
+static const tran_high_t cospi_17_64 = 11003;
+static const tran_high_t cospi_18_64 = 10394;
+static const tran_high_t cospi_19_64 = 9760;
+static const tran_high_t cospi_20_64 = 9102;
+static const tran_high_t cospi_21_64 = 8423;
+static const tran_high_t cospi_22_64 = 7723;
+static const tran_high_t cospi_23_64 = 7005;
+static const tran_high_t cospi_24_64 = 6270;
+static const tran_high_t cospi_25_64 = 5520;
+static const tran_high_t cospi_26_64 = 4756;
+static const tran_high_t cospi_27_64 = 3981;
+static const tran_high_t cospi_28_64 = 3196;
+static const tran_high_t cospi_29_64 = 2404;
+static const tran_high_t cospi_30_64 = 1606;
+static const tran_high_t cospi_31_64 = 804;
+
+//  16384 * sqrt(2) * sin(kPi/9) * 2 / 3
+static const tran_high_t sinpi_1_9 = 5283;
+static const tran_high_t sinpi_2_9 = 9929;
+static const tran_high_t sinpi_3_9 = 13377;
+static const tran_high_t sinpi_4_9 = 15212;
+
+static INLINE tran_low_t check_range(tran_high_t input) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid VP9 input streams, intermediate stage coefficients should always
+  // stay within the range of a signed 16 bit integer. Coefficients can go out
+  // of this range for invalid/corrupt VP9 streams. However, strictly checking
+  // this range for every intermediate coefficient can burdensome for a decoder,
+  // therefore the following assertion is only enabled when configured with
+  // --enable-coefficient-range-checking.
+  assert(INT16_MIN <= input);
+  assert(input <= INT16_MAX);
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  return (tran_low_t)input;
+}
+
+static INLINE tran_low_t dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return check_range(rv);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE tran_low_t highbd_check_range(tran_high_t input,
+                                            int bd) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid highbitdepth VP9 streams, intermediate stage coefficients will
+  // stay within the ranges:
+  // - 8 bit: signed 16 bit integer
+  // - 10 bit: signed 18 bit integer
+  // - 12 bit: signed 20 bit integer
+  const int32_t int_max = (1 << (7 + bd)) - 1;
+  const int32_t int_min = -int_max - 1;
+  assert(int_min <= input);
+  assert(input <= int_max);
+  (void) int_min;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  (void) bd;
+  return (tran_low_t)input;
+}
+
+static INLINE tran_low_t highbd_dct_const_round_shift(tran_high_t input,
+                                                      int bd) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return highbd_check_range(rv, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+typedef void (*transform_1d)(const tran_low_t*, tran_low_t*);
+
+typedef struct {
+  transform_1d cols, rows;  // vertical and horizontal
+} transform_2d;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*highbd_transform_1d)(const tran_low_t*, tran_low_t*, int bd);
+
+typedef struct {
+  highbd_transform_1d cols, rows;  // vertical and horizontal
+} highbd_transform_2d;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#if CONFIG_EMULATE_HARDWARE
+// When CONFIG_EMULATE_HARDWARE is 1 the transform performs a
+// non-normative method to handle overflows. A stream that causes
+// overflows  in the inverse transform is considered invalid in VP9,
+// and a hardware implementer is free to choose any reasonable
+// method to handle overflows. However to aid in hardware
+// verification they can use a specific implementation of the
+// WRAPLOW() macro below that is identical to their intended
+// hardware implementation (and also use configure options to trigger
+// the C-implementation of the transform).
+//
+// The particular WRAPLOW implementation below performs strict
+// overflow wrapping to match common hardware implementations.
+// bd of 8 uses trans_low with 16bits, need to remove 16bits
+// bd of 10 uses trans_low with 18bits, need to remove 14bits
+// bd of 12 uses trans_low with 20bits, need to remove 12bits
+// bd of x uses trans_low with 8+x bits, need to remove 24-x bits
+#define WRAPLOW(x, bd) ((((int32_t)(x)) << (24 - bd)) >> (24 - bd))
+#else
+#define WRAPLOW(x, bd) (x)
+#endif  // CONFIG_EMULATE_HARDWARE
+
+void vp9_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void vp9_idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride, int
+                       eob);
+void vp9_idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                       int eob);
+
+void vp9_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob);
+void vp9_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                    int stride, int eob);
+void vp9_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input, uint8_t *dest,
+                      int stride, int eob);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_idct4(const tran_low_t *input, tran_low_t *output, int bd);
+void vp9_highbd_idct8(const tran_low_t *input, tran_low_t *output, int bd);
+void vp9_highbd_idct16(const tran_low_t *input, tran_low_t *output, int bd);
+void vp9_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void vp9_highbd_idct16x16_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd);
+void vp9_highbd_idct32x32_add(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, int bd);
+void vp9_highbd_iht4x4_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd);
+void vp9_highbd_iht8x8_add(TX_TYPE tx_type, const tran_low_t *input,
+                           uint8_t *dest, int stride, int eob, int bd);
+void vp9_highbd_iht16x16_add(TX_TYPE tx_type, const tran_low_t *input,
+                             uint8_t *dest, int stride, int eob, int bd);
+static INLINE uint16_t highbd_clip_pixel_add(uint16_t dest, tran_high_t trans,
+                                             int bd) {
+  trans = WRAPLOW(trans, bd);
+  return clip_pixel_highbd(WRAPLOW(dest + trans, bd), bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_IDCT_H_
diff --git a/media/libvpx/vp9/common/vp9_loopfilter.c b/media/libvpx/vp9/common/vp9_loopfilter.c
new file mode 100644
index 000000000..484e457df
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_loopfilter.c
@@ -0,0 +1,1615 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_seg_common.h"
+
+// 64 bit masks for left transform size. Each 1 represents a position where
+// we should apply a loop filter across the left border of an 8x8 block
+// boundary.
+//
+// In the case of TX_16X16->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//
+// A loopfilter should be applied to every other 8x8 horizontally.
+static const uint64_t left_64x64_txform_mask[TX_SIZES]= {
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x5555555555555555ULL,  // TX_16x16
+  0x1111111111111111ULL,  // TX_32x32
+};
+
+// 64 bit masks for above transform size. Each 1 represents a position where
+// we should apply a loop filter across the top border of an 8x8 block
+// boundary.
+//
+// In the case of TX_32x32 ->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//
+// A loopfilter should be applied to every other 4 the row vertically.
+static const uint64_t above_64x64_txform_mask[TX_SIZES]= {
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x00ff00ff00ff00ffULL,  // TX_16x16
+  0x000000ff000000ffULL,  // TX_32x32
+};
+
+// 64 bit masks for prediction sizes (left). Each 1 represents a position
+// where left border of an 8x8 block. These are aligned to the right most
+// appropriate bit, and then shifted into place.
+//
+// In the case of TX_16x32 ->  ( low order byte first ) we end up with
+// a mask that looks like this :
+//
+//  10000000
+//  10000000
+//  10000000
+//  10000000
+//  00000000
+//  00000000
+//  00000000
+//  00000000
+static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4,
+  0x0000000000000001ULL,  // BLOCK_4X8,
+  0x0000000000000001ULL,  // BLOCK_8X4,
+  0x0000000000000001ULL,  // BLOCK_8X8,
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000001ULL,  // BLOCK_16X8,
+  0x0000000000000101ULL,  // BLOCK_16X16,
+  0x0000000001010101ULL,  // BLOCK_16X32,
+  0x0000000000000101ULL,  // BLOCK_32X16,
+  0x0000000001010101ULL,  // BLOCK_32X32,
+  0x0101010101010101ULL,  // BLOCK_32X64,
+  0x0000000001010101ULL,  // BLOCK_64X32,
+  0x0101010101010101ULL,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each prediction size.
+static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000001ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000003ULL,  // BLOCK_16X16
+  0x0000000000000003ULL,  // BLOCK_16X32,
+  0x000000000000000fULL,  // BLOCK_32X16,
+  0x000000000000000fULL,  // BLOCK_32X32,
+  0x000000000000000fULL,  // BLOCK_32X64,
+  0x00000000000000ffULL,  // BLOCK_64X32,
+  0x00000000000000ffULL,  // BLOCK_64X64
+};
+// 64 bit mask to shift and set for each prediction size. A bit is set for
+// each 8x8 block that would be in the left most block of the given block
+// size in the 64x64 block.
+static const uint64_t size_mask[BLOCK_SIZES] = {
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000303ULL,  // BLOCK_16X16
+  0x0000000003030303ULL,  // BLOCK_16X32,
+  0x0000000000000f0fULL,  // BLOCK_32X16,
+  0x000000000f0f0f0fULL,  // BLOCK_32X32,
+  0x0f0f0f0f0f0f0f0fULL,  // BLOCK_32X64,
+  0x00000000ffffffffULL,  // BLOCK_64X32,
+  0xffffffffffffffffULL,  // BLOCK_64X64
+};
+
+// These are used for masking the left and above borders.
+static const uint64_t left_border =  0x1111111111111111ULL;
+static const uint64_t above_border = 0x000000ff000000ffULL;
+
+// 16 bit masks for uv transform sizes.
+static const uint16_t left_64x64_txform_mask_uv[TX_SIZES]= {
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x5555,  // TX_16x16
+  0x1111,  // TX_32x32
+};
+
+static const uint16_t above_64x64_txform_mask_uv[TX_SIZES]= {
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x0f0f,  // TX_16x16
+  0x000f,  // TX_32x32
+};
+
+// 16 bit left mask to shift and set for each uv prediction size.
+static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4,
+  0x0001,  // BLOCK_4X8,
+  0x0001,  // BLOCK_8X4,
+  0x0001,  // BLOCK_8X8,
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8,
+  0x0001,  // BLOCK_16X16,
+  0x0011,  // BLOCK_16X32,
+  0x0001,  // BLOCK_32X16,
+  0x0011,  // BLOCK_32X32,
+  0x1111,  // BLOCK_32X64
+  0x0011,  // BLOCK_64X32,
+  0x1111,  // BLOCK_64X64
+};
+// 16 bit above mask to shift and set for uv each prediction size.
+static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0001,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0003,  // BLOCK_32X32,
+  0x0003,  // BLOCK_32X64,
+  0x000f,  // BLOCK_64X32,
+  0x000f,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each uv prediction size
+static const uint16_t size_mask_uv[BLOCK_SIZES] = {
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0011,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0033,  // BLOCK_32X32,
+  0x3333,  // BLOCK_32X64,
+  0x00ff,  // BLOCK_64X32,
+  0xffff,  // BLOCK_64X64
+};
+static const uint16_t left_border_uv =  0x1111;
+static const uint16_t above_border_uv = 0x000f;
+
+static const int mode_lf_lut[MB_MODE_COUNT] = {
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
+  1, 1, 0, 1                     // INTER_MODES (ZEROMV == 0)
+};
+
+static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
+  int lvl;
+
+  // For each possible value for the loop filter fill out limits
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
+    // Set loop filter parameters that control sharpness.
+    int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
+
+    if (sharpness_lvl > 0) {
+      if (block_inside_limit > (9 - sharpness_lvl))
+        block_inside_limit = (9 - sharpness_lvl);
+    }
+
+    if (block_inside_limit < 1)
+      block_inside_limit = 1;
+
+    memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
+    memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
+           SIMD_WIDTH);
+  }
+}
+
+static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
+                                const MB_MODE_INFO *mbmi) {
+  return lfi_n->lvl[mbmi->segment_id][mbmi->ref_frame[0]]
+                   [mode_lf_lut[mbmi->mode]];
+}
+
+void vp9_loop_filter_init(VP9_COMMON *cm) {
+  loop_filter_info_n *lfi = &cm->lf_info;
+  struct loopfilter *lf = &cm->lf;
+  int lvl;
+
+  // init limits for given sharpness
+  update_sharpness(lfi, lf->sharpness_level);
+  lf->last_sharpness_level = lf->sharpness_level;
+
+  // init hev threshold const vectors
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
+    memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
+}
+
+void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
+  int seg_id;
+  // n_shift is the multiplier for lf_deltas
+  // the multiplier is 1 for when filter_lvl is between 0 and 31;
+  // 2 when filter_lvl is between 32 and 63
+  const int scale = 1 << (default_filt_lvl >> 5);
+  loop_filter_info_n *const lfi = &cm->lf_info;
+  struct loopfilter *const lf = &cm->lf;
+  const struct segmentation *const seg = &cm->seg;
+
+  // update limits if sharpness has changed
+  if (lf->last_sharpness_level != lf->sharpness_level) {
+    update_sharpness(lfi, lf->sharpness_level);
+    lf->last_sharpness_level = lf->sharpness_level;
+  }
+
+  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
+    int lvl_seg = default_filt_lvl;
+    if (vp9_segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
+      const int data = vp9_get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
+      lvl_seg = clamp(seg->abs_delta == SEGMENT_ABSDATA ?
+                      data : default_filt_lvl + data,
+                      0, MAX_LOOP_FILTER);
+    }
+
+    if (!lf->mode_ref_delta_enabled) {
+      // we could get rid of this if we assume that deltas are set to
+      // zero when not in use; encoder always uses deltas
+      memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
+    } else {
+      int ref, mode;
+      const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
+      lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
+
+      for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) {
+        for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
+          const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale
+                                        + lf->mode_deltas[mode] * scale;
+          lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
+        }
+      }
+    }
+  }
+}
+
+static void filter_selectively_vert_row2(int subsampling_factor,
+                                         uint8_t *s, int pitch,
+                                         unsigned int mask_16x16_l,
+                                         unsigned int mask_8x8_l,
+                                         unsigned int mask_4x4_l,
+                                         unsigned int mask_4x4_int_l,
+                                         const loop_filter_info_n *lfi_n,
+                                         const uint8_t *lfl) {
+  const int mask_shift = subsampling_factor ? 4 : 8;
+  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+
+  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
+  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
+  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
+  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
+  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
+  unsigned int mask;
+
+  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
+              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
+    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
+
+    // TODO(yunqingwang): count in loopfilter functions should be removed.
+    if (mask & 1) {
+      if ((mask_16x16_0 | mask_16x16_1) & 1) {
+        if ((mask_16x16_0 & mask_16x16_1) & 1) {
+          vp9_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                   lfi0->hev_thr);
+        } else if (mask_16x16_0 & 1) {
+          vp9_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
+                              lfi0->hev_thr);
+        } else {
+          vp9_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
+                              lfi1->lim, lfi1->hev_thr);
+        }
+      }
+
+      if ((mask_8x8_0 | mask_8x8_1) & 1) {
+        if ((mask_8x8_0 & mask_8x8_1) & 1) {
+          vp9_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_8x8_0 & 1) {
+          vp9_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
+                             1);
+        } else {
+          vp9_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr, 1);
+        }
+      }
+
+      if ((mask_4x4_0 | mask_4x4_1) & 1) {
+        if ((mask_4x4_0 & mask_4x4_1) & 1) {
+          vp9_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_4x4_0 & 1) {
+          vp9_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr,
+                             1);
+        } else {
+          vp9_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr, 1);
+        }
+      }
+
+      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
+        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
+          vp9_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_4x4_int_0 & 1) {
+          vp9_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                             lfi0->hev_thr, 1);
+        } else {
+          vp9_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr, 1);
+        }
+      }
+    }
+
+    s += 8;
+    lfl += 1;
+    mask_16x16_0 >>= 1;
+    mask_8x8_0 >>= 1;
+    mask_4x4_0 >>= 1;
+    mask_4x4_int_0 >>= 1;
+    mask_16x16_1 >>= 1;
+    mask_8x8_1 >>= 1;
+    mask_4x4_1 >>= 1;
+    mask_4x4_int_1 >>= 1;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_vert_row2(int subsampling_factor,
+                                                uint16_t *s, int pitch,
+                                                unsigned int mask_16x16_l,
+                                                unsigned int mask_8x8_l,
+                                                unsigned int mask_4x4_l,
+                                                unsigned int mask_4x4_int_l,
+                                                const loop_filter_info_n *lfi_n,
+                                                const uint8_t *lfl, int bd) {
+  const int mask_shift = subsampling_factor ? 4 : 8;
+  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+
+  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
+  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
+  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
+  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
+  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
+  unsigned int mask;
+
+  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
+       mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
+    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
+
+    // TODO(yunqingwang): count in loopfilter functions should be removed.
+    if (mask & 1) {
+      if ((mask_16x16_0 | mask_16x16_1) & 1) {
+        if ((mask_16x16_0 & mask_16x16_1) & 1) {
+          vp9_highbd_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                          lfi0->hev_thr, bd);
+        } else if (mask_16x16_0 & 1) {
+          vp9_highbd_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
+                                     lfi0->hev_thr, bd);
+        } else {
+          vp9_highbd_lpf_vertical_16(s + 8 *pitch, pitch, lfi1->mblim,
+                                     lfi1->lim, lfi1->hev_thr, bd);
+        }
+      }
+
+      if ((mask_8x8_0 | mask_8x8_1) & 1) {
+        if ((mask_8x8_0 & mask_8x8_1) & 1) {
+          vp9_highbd_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_8x8_0 & 1) {
+          vp9_highbd_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, 1, bd);
+        } else {
+          vp9_highbd_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, 1, bd);
+        }
+      }
+
+      if ((mask_4x4_0 | mask_4x4_1) & 1) {
+        if ((mask_4x4_0 & mask_4x4_1) & 1) {
+          vp9_highbd_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_4x4_0 & 1) {
+          vp9_highbd_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, 1, bd);
+        } else {
+          vp9_highbd_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, 1, bd);
+        }
+      }
+
+      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
+        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
+          vp9_highbd_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_4x4_int_0 & 1) {
+          vp9_highbd_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, 1, bd);
+        } else {
+          vp9_highbd_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, 1, bd);
+        }
+      }
+    }
+
+    s += 8;
+    lfl += 1;
+    mask_16x16_0 >>= 1;
+    mask_8x8_0 >>= 1;
+    mask_4x4_0 >>= 1;
+    mask_4x4_int_0 >>= 1;
+    mask_16x16_1 >>= 1;
+    mask_8x8_1 >>= 1;
+    mask_4x4_1 >>= 1;
+    mask_4x4_int_1 >>= 1;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void filter_selectively_horiz(uint8_t *s, int pitch,
+                                     unsigned int mask_16x16,
+                                     unsigned int mask_8x8,
+                                     unsigned int mask_4x4,
+                                     unsigned int mask_4x4_int,
+                                     const loop_filter_info_n *lfi_n,
+                                     const uint8_t *lfl) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= count) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    count = 1;
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, 2);
+          count = 2;
+        } else {
+          vp9_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, 1);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          vp9_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+
+          if ((mask_4x4_int & 3) == 3) {
+            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr, 1);
+            else if (mask_4x4_int & 2)
+              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr, 1);
+          }
+          count = 2;
+        } else {
+          vp9_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+          if (mask_4x4_int & 1)
+            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr, 1);
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          vp9_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+          if ((mask_4x4_int & 3) == 3) {
+            vp9_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr, 1);
+            else if (mask_4x4_int & 2)
+              vp9_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr, 1);
+          }
+          count = 2;
+        } else {
+          vp9_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+
+          if (mask_4x4_int & 1)
+            vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr, 1);
+        }
+      } else if (mask_4x4_int & 1) {
+        vp9_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                             lfi->hev_thr, 1);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_horiz(uint16_t *s, int pitch,
+                                            unsigned int mask_16x16,
+                                            unsigned int mask_8x8,
+                                            unsigned int mask_4x4,
+                                            unsigned int mask_4x4_int,
+                                            const loop_filter_info_n *lfi_n,
+                                            const uint8_t *lfl, int bd) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= count) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    count = 1;
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          vp9_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
+                                       lfi->hev_thr, 2, bd);
+          count = 2;
+        } else {
+          vp9_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
+                                       lfi->hev_thr, 1, bd);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          vp9_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+
+          if ((mask_4x4_int & 3) == 3) {
+            vp9_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                             lfi->lim, lfi->hev_thr,
+                                             lfin->mblim, lfin->lim,
+                                             lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              vp9_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, 1, bd);
+            } else if (mask_4x4_int & 2) {
+              vp9_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, 1, bd);
+            }
+          }
+          count = 2;
+        } else {
+          vp9_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, 1, bd);
+
+          if (mask_4x4_int & 1) {
+            vp9_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, 1, bd);
+          }
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          vp9_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+          if ((mask_4x4_int & 3) == 3) {
+            vp9_highbd_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                             lfi->lim, lfi->hev_thr,
+                                             lfin->mblim, lfin->lim,
+                                             lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              vp9_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, 1, bd);
+            } else if (mask_4x4_int & 2) {
+              vp9_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, 1, bd);
+            }
+          }
+          count = 2;
+        } else {
+          vp9_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, 1, bd);
+
+          if (mask_4x4_int & 1) {
+            vp9_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, 1, bd);
+          }
+        }
+      } else if (mask_4x4_int & 1) {
+        vp9_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, 1, bd);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+// This function ors into the current lfm structure, where to do loop
+// filters for the specific mi we are looking at. It uses information
+// including the block_size_type (32x16, 32x32, etc.), the transform size,
+// whether there were any coefficients encoded, and the loop filter strength
+// block we are currently looking at. Shift is used to position the
+// 1's we produce.
+// TODO(JBB) Need another function for different resolution color..
+static void build_masks(const loop_filter_info_n *const lfi_n,
+                        const MODE_INFO *mi, const int shift_y,
+                        const int shift_uv,
+                        LOOP_FILTER_MASK *lfm) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+  const BLOCK_SIZE block_size = mbmi->sb_type;
+  const TX_SIZE tx_size_y = mbmi->tx_size;
+  const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1);
+  const int filter_level = get_filter_level(lfi_n, mbmi);
+  uint64_t *const left_y = &lfm->left_y[tx_size_y];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
+  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
+  uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
+  int i;
+
+  // If filter level is 0 we don't loop filter.
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    int index = shift_y;
+    for (i = 0; i < h; i++) {
+      memset(&lfm->lfl_y[index], filter_level, w);
+      index += 8;
+    }
+  }
+
+  // These set 1 in the current block size for the block size edges.
+  // For instance if the block size is 32x16, we'll set:
+  //    above =   1111
+  //              0000
+  //    and
+  //    left  =   1000
+  //          =   1000
+  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
+  //        1,  not 8...
+  //
+  // U and V set things on a 16 bit scale.
+  //
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+  *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
+
+  // If the block has no coefficients and is not intra we skip applying
+  // the loop filter on block edges.
+  if (mbmi->skip && is_inter_block(mbmi))
+    return;
+
+  // Here we are adding a mask for the transform size. The transform
+  // size mask is set to be correct for a 64x64 prediction block size. We
+  // mask to match the size of the block we are working on and then shift it
+  // into place..
+  *above_y |= (size_mask[block_size] &
+               above_64x64_txform_mask[tx_size_y]) << shift_y;
+  *above_uv |= (size_mask_uv[block_size] &
+                above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+
+  *left_y |= (size_mask[block_size] &
+              left_64x64_txform_mask[tx_size_y]) << shift_y;
+  *left_uv |= (size_mask_uv[block_size] &
+               left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv;
+
+  // Here we are trying to determine what to do with the internal 4x4 block
+  // boundaries.  These differ from the 4x4 boundaries on the outside edge of
+  // an 8x8 in that the internal ones can be skipped and don't depend on
+  // the prediction block size.
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
+
+  if (tx_size_uv == TX_4X4)
+    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
+}
+
+// This function does the same thing as the one above with the exception that
+// it only affects the y masks. It exists because for blocks < 16x16 in size,
+// we only update u and v masks on the first block.
+static void build_y_mask(const loop_filter_info_n *const lfi_n,
+                         const MODE_INFO *mi, const int shift_y,
+                         LOOP_FILTER_MASK *lfm) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+  const BLOCK_SIZE block_size = mbmi->sb_type;
+  const TX_SIZE tx_size_y = mbmi->tx_size;
+  const int filter_level = get_filter_level(lfi_n, mbmi);
+  uint64_t *const left_y = &lfm->left_y[tx_size_y];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  int i;
+
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    int index = shift_y;
+    for (i = 0; i < h; i++) {
+      memset(&lfm->lfl_y[index], filter_level, w);
+      index += 8;
+    }
+  }
+
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+
+  if (mbmi->skip && is_inter_block(mbmi))
+    return;
+
+  *above_y |= (size_mask[block_size] &
+               above_64x64_txform_mask[tx_size_y]) << shift_y;
+
+  *left_y |= (size_mask[block_size] &
+              left_64x64_txform_mask[tx_size_y]) << shift_y;
+
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
+}
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+// TODO(JBB): This function only works for yv12.
+void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
+                    MODE_INFO **mi, const int mode_info_stride,
+                    LOOP_FILTER_MASK *lfm) {
+  int idx_32, idx_16, idx_8;
+  const loop_filter_info_n *const lfi_n = &cm->lf_info;
+  MODE_INFO **mip = mi;
+  MODE_INFO **mip2 = mi;
+
+  // These are offsets to the next mi in the 64x64 block. It is what gets
+  // added to the mi ptr as we go through each loop. It helps us to avoid
+  // setting up special row and column counters for each index. The last step
+  // brings us out back to the starting position.
+  const int offset_32[] = {4, (mode_info_stride << 2) - 4, 4,
+                           -(mode_info_stride << 2) - 4};
+  const int offset_16[] = {2, (mode_info_stride << 1) - 2, 2,
+                           -(mode_info_stride << 1) - 2};
+  const int offset[] = {1, mode_info_stride - 1, 1, -mode_info_stride - 1};
+
+  // Following variables represent shifts to position the current block
+  // mask over the appropriate block. A shift of 36 to the left will move
+  // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
+  // 4 rows to the appropriate spot.
+  const int shift_32_y[] = {0, 4, 32, 36};
+  const int shift_16_y[] = {0, 2, 16, 18};
+  const int shift_8_y[] = {0, 1, 8, 9};
+  const int shift_32_uv[] = {0, 2, 8, 10};
+  const int shift_16_uv[] = {0, 1, 4, 5};
+  int i;
+  const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ?
+                        cm->mi_rows - mi_row : MI_BLOCK_SIZE);
+  const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ?
+                        cm->mi_cols - mi_col : MI_BLOCK_SIZE);
+
+  vp9_zero(*lfm);
+  assert(mip[0] != NULL);
+
+  // TODO(jimbankoski): Try moving most of the following code into decode
+  // loop and storing lfm in the mbmi structure so that we don't have to go
+  // through the recursive loop structure multiple times.
+  switch (mip[0]->mbmi.sb_type) {
+    case BLOCK_64X64:
+      build_masks(lfi_n, mip[0] , 0, 0, lfm);
+      break;
+    case BLOCK_64X32:
+      build_masks(lfi_n, mip[0], 0, 0, lfm);
+      mip2 = mip + mode_info_stride * 4;
+      if (4 >= max_rows)
+        break;
+      build_masks(lfi_n, mip2[0], 32, 8, lfm);
+      break;
+    case BLOCK_32X64:
+      build_masks(lfi_n, mip[0], 0, 0, lfm);
+      mip2 = mip + 4;
+      if (4 >= max_cols)
+        break;
+      build_masks(lfi_n, mip2[0], 4, 2, lfm);
+      break;
+    default:
+      for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
+        const int shift_y = shift_32_y[idx_32];
+        const int shift_uv = shift_32_uv[idx_32];
+        const int mi_32_col_offset = ((idx_32 & 1) << 2);
+        const int mi_32_row_offset = ((idx_32 >> 1) << 2);
+        if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
+          continue;
+        switch (mip[0]->mbmi.sb_type) {
+          case BLOCK_32X32:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            break;
+          case BLOCK_32X16:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            if (mi_32_row_offset + 2 >= max_rows)
+              continue;
+            mip2 = mip + mode_info_stride * 2;
+            build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
+            break;
+          case BLOCK_16X32:
+            build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+            if (mi_32_col_offset + 2 >= max_cols)
+              continue;
+            mip2 = mip + 2;
+            build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
+            break;
+          default:
+            for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
+              const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
+              const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
+              const int mi_16_col_offset = mi_32_col_offset +
+                  ((idx_16 & 1) << 1);
+              const int mi_16_row_offset = mi_32_row_offset +
+                  ((idx_16 >> 1) << 1);
+
+              if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
+                continue;
+
+              switch (mip[0]->mbmi.sb_type) {
+                case BLOCK_16X16:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  break;
+                case BLOCK_16X8:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  if (mi_16_row_offset + 1 >= max_rows)
+                    continue;
+                  mip2 = mip + mode_info_stride;
+                  build_y_mask(lfi_n, mip2[0], shift_y+8, lfm);
+                  break;
+                case BLOCK_8X16:
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  if (mi_16_col_offset +1 >= max_cols)
+                    continue;
+                  mip2 = mip + 1;
+                  build_y_mask(lfi_n, mip2[0], shift_y+1, lfm);
+                  break;
+                default: {
+                  const int shift_y = shift_32_y[idx_32] +
+                                      shift_16_y[idx_16] +
+                                      shift_8_y[0];
+                  build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
+                  mip += offset[0];
+                  for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
+                    const int shift_y = shift_32_y[idx_32] +
+                                        shift_16_y[idx_16] +
+                                        shift_8_y[idx_8];
+                    const int mi_8_col_offset = mi_16_col_offset +
+                        ((idx_8 & 1));
+                    const int mi_8_row_offset = mi_16_row_offset +
+                        ((idx_8 >> 1));
+
+                    if (mi_8_col_offset >= max_cols ||
+                        mi_8_row_offset >= max_rows)
+                      continue;
+                    build_y_mask(lfi_n, mip[0], shift_y, lfm);
+                  }
+                  break;
+                }
+              }
+            }
+            break;
+        }
+      }
+      break;
+  }
+  // The largest loopfilter we have is 16x16 so we use the 16x16 mask
+  // for 32x32 transforms also.
+  lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
+  lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
+  lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
+  lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
+
+  // We do at least 8 tap filter on every 32x32 even if the transform size
+  // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
+  // remove it from the 4x4.
+  lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
+  lfm->left_y[TX_4X4] &= ~left_border;
+  lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
+  lfm->above_y[TX_4X4] &= ~above_border;
+  lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
+  lfm->left_uv[TX_4X4] &= ~left_border_uv;
+  lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
+  lfm->above_uv[TX_4X4] &= ~above_border_uv;
+
+  // We do some special edge handling.
+  if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
+    const uint64_t rows = cm->mi_rows - mi_row;
+
+    // Each pixel inside the border gets a 1,
+    const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1);
+    const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1);
+
+    // Remove values completely outside our border.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->int_4x4_uv &= mask_uv;
+
+    // We don't apply a wide loop filter on the last uv block row. If set
+    // apply the shorter one instead.
+    if (rows == 1) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
+      lfm->above_uv[TX_16X16] = 0;
+    }
+    if (rows == 5) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
+      lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
+    }
+  }
+
+  if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
+    const uint64_t columns = cm->mi_cols - mi_col;
+
+    // Each pixel inside the border gets a 1, the multiply copies the border
+    // to where we need it.
+    const uint64_t mask_y  = (((1 << columns) - 1)) * 0x0101010101010101ULL;
+    const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
+
+    // Internal edges are not applied on the last column of the image so
+    // we mask 1 more for the internal edges
+    const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
+
+    // Remove the bits outside the image edge.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->int_4x4_uv &= mask_uv_int;
+
+    // We don't apply a wide loop filter on the last uv column. If set
+    // apply the shorter one instead.
+    if (columns == 1) {
+      lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
+      lfm->left_uv[TX_16X16] = 0;
+    }
+    if (columns == 5) {
+      lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
+      lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
+    }
+  }
+  // We don't apply a loop filter on the first column in the image, mask that
+  // out.
+  if (mi_col == 0) {
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= 0xfefefefefefefefeULL;
+      lfm->left_uv[i] &= 0xeeee;
+    }
+  }
+
+  // Assert if we try to apply 2 different loop filters at the same position.
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
+  assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8]));
+  assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
+}
+
+static void filter_selectively_vert(uint8_t *s, int pitch,
+                                    unsigned int mask_16x16,
+                                    unsigned int mask_8x8,
+                                    unsigned int mask_4x4,
+                                    unsigned int mask_4x4_int,
+                                    const loop_filter_info_n *lfi_n,
+                                    const uint8_t *lfl) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        vp9_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      } else if (mask_8x8 & 1) {
+        vp9_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+      } else if (mask_4x4 & 1) {
+        vp9_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+      }
+    }
+    if (mask_4x4_int & 1)
+      vp9_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr, 1);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void highbd_filter_selectively_vert(uint16_t *s, int pitch,
+                                           unsigned int mask_16x16,
+                                           unsigned int mask_8x8,
+                                           unsigned int mask_4x4,
+                                           unsigned int mask_4x4_int,
+                                           const loop_filter_info_n *lfi_n,
+                                           const uint8_t *lfl, int bd) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        vp9_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr, bd);
+      } else if (mask_8x8 & 1) {
+        vp9_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim,
+                                  lfi->hev_thr, 1, bd);
+      } else if (mask_4x4 & 1) {
+        vp9_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, 1, bd);
+      }
+    }
+    if (mask_4x4_int & 1)
+      vp9_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, 1, bd);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_filter_block_plane_non420(VP9_COMMON *cm,
+                                   struct macroblockd_plane *plane,
+                                   MODE_INFO **mi_8x8,
+                                   int mi_row, int mi_col) {
+  const int ss_x = plane->subsampling_x;
+  const int ss_y = plane->subsampling_y;
+  const int row_step = 1 << ss_y;
+  const int col_step = 1 << ss_x;
+  const int row_step_stride = cm->mi_stride * row_step;
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t* const dst0 = dst->buf;
+  unsigned int mask_16x16[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_8x8[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_4x4[MI_BLOCK_SIZE] = {0};
+  unsigned int mask_4x4_int[MI_BLOCK_SIZE] = {0};
+  uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
+  int r, c;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
+    unsigned int mask_16x16_c = 0;
+    unsigned int mask_8x8_c = 0;
+    unsigned int mask_4x4_c = 0;
+    unsigned int border_mask;
+
+    // Determine the vertical edges that need filtering
+    for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
+      const MODE_INFO *mi = mi_8x8[c];
+      const BLOCK_SIZE sb_type = mi[0].mbmi.sb_type;
+      const int skip_this = mi[0].mbmi.skip && is_inter_block(&mi[0].mbmi);
+      // left edge of current unit is block/partition edge -> no skip
+      const int block_edge_left = (num_4x4_blocks_wide_lookup[sb_type] > 1) ?
+          !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1)) : 1;
+      const int skip_this_c = skip_this && !block_edge_left;
+      // top edge of current unit is block/partition edge -> no skip
+      const int block_edge_above = (num_4x4_blocks_high_lookup[sb_type] > 1) ?
+          !(r & (num_8x8_blocks_high_lookup[sb_type] - 1)) : 1;
+      const int skip_this_r = skip_this && !block_edge_above;
+      const TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
+                            ? get_uv_tx_size(&mi[0].mbmi, plane)
+                            : mi[0].mbmi.tx_size;
+      const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
+      const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
+
+      // Filter level can vary per MI
+      if (!(lfl[(r << 3) + (c >> ss_x)] =
+            get_filter_level(&cm->lf_info, &mi[0].mbmi)))
+        continue;
+
+      // Build masks based on the transform size of each block
+      if (tx_size == TX_32X32) {
+        if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
+          if (!skip_border_4x4_c)
+            mask_16x16_c |= 1 << (c >> ss_x);
+          else
+            mask_8x8_c |= 1 << (c >> ss_x);
+        }
+        if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
+          if (!skip_border_4x4_r)
+            mask_16x16[r] |= 1 << (c >> ss_x);
+          else
+            mask_8x8[r] |= 1 << (c >> ss_x);
+        }
+      } else if (tx_size == TX_16X16) {
+        if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
+          if (!skip_border_4x4_c)
+            mask_16x16_c |= 1 << (c >> ss_x);
+          else
+            mask_8x8_c |= 1 << (c >> ss_x);
+        }
+        if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
+          if (!skip_border_4x4_r)
+            mask_16x16[r] |= 1 << (c >> ss_x);
+          else
+            mask_8x8[r] |= 1 << (c >> ss_x);
+        }
+      } else {
+        // force 8x8 filtering on 32x32 boundaries
+        if (!skip_this_c) {
+          if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
+            mask_8x8_c |= 1 << (c >> ss_x);
+          else
+            mask_4x4_c |= 1 << (c >> ss_x);
+        }
+
+        if (!skip_this_r) {
+          if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
+            mask_8x8[r] |= 1 << (c >> ss_x);
+          else
+            mask_4x4[r] |= 1 << (c >> ss_x);
+        }
+
+        if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
+          mask_4x4_int[r] |= 1 << (c >> ss_x);
+      }
+    }
+
+    // Disable filtering on the leftmost column
+    border_mask = ~(mi_col == 0);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_vert(CONVERT_TO_SHORTPTR(dst->buf),
+                                     dst->stride,
+                                     mask_16x16_c & border_mask,
+                                     mask_8x8_c & border_mask,
+                                     mask_4x4_c & border_mask,
+                                     mask_4x4_int[r],
+                                     &cm->lf_info, &lfl[r << 3],
+                                     (int)cm->bit_depth);
+    } else {
+      filter_selectively_vert(dst->buf, dst->stride,
+                              mask_16x16_c & border_mask,
+                              mask_8x8_c & border_mask,
+                              mask_4x4_c & border_mask,
+                              mask_4x4_int[r],
+                              &cm->lf_info, &lfl[r << 3]);
+    }
+#else
+    filter_selectively_vert(dst->buf, dst->stride,
+                            mask_16x16_c & border_mask,
+                            mask_8x8_c & border_mask,
+                            mask_4x4_c & border_mask,
+                            mask_4x4_int[r],
+                            &cm->lf_info, &lfl[r << 3]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 8 * dst->stride;
+    mi_8x8 += row_step_stride;
+  }
+
+  // Now do horizontal pass
+  dst->buf = dst0;
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
+    const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
+    const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
+
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16[r];
+      mask_8x8_r = mask_8x8[r];
+      mask_4x4_r = mask_4x4[r];
+    }
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
+                                      dst->stride,
+                                      mask_16x16_r,
+                                      mask_8x8_r,
+                                      mask_4x4_r,
+                                      mask_4x4_int_r,
+                                      &cm->lf_info, &lfl[r << 3],
+                                      (int)cm->bit_depth);
+    } else {
+      filter_selectively_horiz(dst->buf, dst->stride,
+                               mask_16x16_r,
+                               mask_8x8_r,
+                               mask_4x4_r,
+                               mask_4x4_int_r,
+                               &cm->lf_info, &lfl[r << 3]);
+    }
+#else
+    filter_selectively_horiz(dst->buf, dst->stride,
+                             mask_16x16_r,
+                             mask_8x8_r,
+                             mask_4x4_r,
+                             mask_4x4_int_r,
+                             &cm->lf_info, &lfl[r << 3]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 8 * dst->stride;
+  }
+}
+
+void vp9_filter_block_plane_ss00(VP9_COMMON *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r;
+  uint64_t mask_16x16 = lfm->left_y[TX_16X16];
+  uint64_t mask_8x8 = lfm->left_y[TX_8X8];
+  uint64_t mask_4x4 = lfm->left_y[TX_4X4];
+  uint64_t mask_4x4_int = lfm->int_4x4_y;
+
+  assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
+    unsigned int mask_16x16_l = mask_16x16 & 0xffff;
+    unsigned int mask_8x8_l = mask_8x8 & 0xffff;
+    unsigned int mask_4x4_l = mask_4x4 & 0xffff;
+    unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
+
+// Disable filtering on the leftmost column.
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_vert_row2(
+          plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
+          mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+          &lfm->lfl_y[r << 3], (int)cm->bit_depth);
+    } else {
+      filter_selectively_vert_row2(
+          plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
+          mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
+    }
+#else
+    filter_selectively_vert_row2(
+        plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
+        mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r << 3]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+    dst->buf += 16 * dst->stride;
+    mask_16x16 >>= 16;
+    mask_8x8 >>= 16;
+    mask_4x4 >>= 16;
+    mask_4x4_int >>= 16;
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+  mask_16x16 = lfm->above_y[TX_16X16];
+  mask_8x8 = lfm->above_y[TX_8X8];
+  mask_4x4 = lfm->above_y[TX_4X4];
+  mask_4x4_int = lfm->int_4x4_y;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xff;
+      mask_8x8_r = mask_8x8 & 0xff;
+      mask_4x4_r = mask_4x4 & 0xff;
+    }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(
+          CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
+          mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info, &lfm->lfl_y[r << 3],
+          (int)cm->bit_depth);
+    } else {
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
+                               &lfm->lfl_y[r << 3]);
+    }
+#else
+    filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                             mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
+                             &lfm->lfl_y[r << 3]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    dst->buf += 8 * dst->stride;
+    mask_16x16 >>= 8;
+    mask_8x8 >>= 8;
+    mask_4x4 >>= 8;
+    mask_4x4_int >>= 8;
+  }
+}
+
+void vp9_filter_block_plane_ss11(VP9_COMMON *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r, c;
+
+  uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
+  uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
+  uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
+  uint16_t mask_4x4_int = lfm->int_4x4_uv;
+
+  assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
+    if (plane->plane_type == 1) {
+      for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
+        lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
+        lfm->lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
+      }
+    }
+
+    {
+      unsigned int mask_16x16_l = mask_16x16 & 0xff;
+      unsigned int mask_8x8_l = mask_8x8 & 0xff;
+      unsigned int mask_4x4_l = mask_4x4 & 0xff;
+      unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;
+
+// Disable filtering on the leftmost column.
+#if CONFIG_VP9_HIGHBITDEPTH
+      if (cm->use_highbitdepth) {
+        highbd_filter_selectively_vert_row2(
+            plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
+            mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+            &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
+      } else {
+        filter_selectively_vert_row2(
+            plane->subsampling_x, dst->buf, dst->stride,
+            mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+            &lfm->lfl_uv[r << 1]);
+      }
+#else
+      filter_selectively_vert_row2(
+          plane->subsampling_x, dst->buf, dst->stride,
+          mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+          &lfm->lfl_uv[r << 1]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+      dst->buf += 16 * dst->stride;
+      mask_16x16 >>= 8;
+      mask_8x8 >>= 8;
+      mask_4x4 >>= 8;
+      mask_4x4_int >>= 8;
+    }
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+  mask_16x16 = lfm->above_uv[TX_16X16];
+  mask_8x8 = lfm->above_uv[TX_8X8];
+  mask_4x4 = lfm->above_uv[TX_4X4];
+  mask_4x4_int = lfm->int_4x4_uv;
+
+  for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
+    const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
+    const unsigned int mask_4x4_int_r =
+        skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xf;
+      mask_8x8_r = mask_8x8 & 0xf;
+      mask_4x4_r = mask_4x4 & 0xf;
+    }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf),
+                                      dst->stride, mask_16x16_r, mask_8x8_r,
+                                      mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
+                                      &lfm->lfl_uv[r << 1], (int)cm->bit_depth);
+    } else {
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
+                               &lfm->lfl_uv[r << 1]);
+    }
+#else
+    filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                             mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
+                             &lfm->lfl_uv[r << 1]);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    dst->buf += 8 * dst->stride;
+    mask_16x16 >>= 4;
+    mask_8x8 >>= 4;
+    mask_4x4 >>= 4;
+    mask_4x4_int >>= 4;
+  }
+}
+
+void vp9_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer,
+                          VP9_COMMON *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
+                          int start, int stop, int y_only) {
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  enum lf_path path;
+  LOOP_FILTER_MASK lfm;
+  int mi_row, mi_col;
+
+  if (y_only)
+    path = LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    path = LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    path = LF_PATH_444;
+  else
+    path = LF_PATH_SLOW;
+
+  for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
+      int plane;
+
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+
+      // TODO(JBB): Make setup_mask work for non 420.
+      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
+                     &lfm);
+
+      vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
+      for (plane = 1; plane < num_planes; ++plane) {
+        switch (path) {
+          case LF_PATH_420:
+            vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_444:
+            vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_SLOW:
+            vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+            break;
+        }
+      }
+    }
+  }
+}
+
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           VP9_COMMON *cm, MACROBLOCKD *xd,
+                           int frame_filter_level,
+                           int y_only, int partial_frame) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+  if (!frame_filter_level) return;
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = MAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  vp9_loop_filter_frame_init(cm, frame_filter_level);
+  vp9_loop_filter_rows(frame, cm, xd->plane,
+                       start_mi_row, end_mi_row,
+                       y_only);
+}
+
+void vp9_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]) {
+  lf_data->frame_buffer = frame_buffer;
+  lf_data->cm = cm;
+  lf_data->start = 0;
+  lf_data->stop = 0;
+  lf_data->y_only = 0;
+  memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
+}
+
+int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
+  (void)unused;
+  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+                       lf_data->start, lf_data->stop, lf_data->y_only);
+  return 1;
+}
diff --git a/media/libvpx/vp9/common/vp9_loopfilter.h b/media/libvpx/vp9/common/vp9_loopfilter.h
new file mode 100644
index 000000000..f7cbde678
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_loopfilter.h
@@ -0,0 +1,155 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_LOOPFILTER_H_
+#define VP9_COMMON_VP9_LOOPFILTER_H_
+
+#include "vpx_ports/mem.h"
+#include "./vpx_config.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LOOP_FILTER 63
+#define MAX_SHARPNESS 7
+
+#define SIMD_WIDTH 16
+
+#define MAX_REF_LF_DELTAS       4
+#define MAX_MODE_LF_DELTAS      2
+
+enum lf_path {
+  LF_PATH_420,
+  LF_PATH_444,
+  LF_PATH_SLOW,
+};
+
+struct loopfilter {
+  int filter_level;
+
+  int sharpness_level;
+  int last_sharpness_level;
+
+  uint8_t mode_ref_delta_enabled;
+  uint8_t mode_ref_delta_update;
+
+  // 0 = Intra, Last, GF, ARF
+  signed char ref_deltas[MAX_REF_LF_DELTAS];
+  signed char last_ref_deltas[MAX_REF_LF_DELTAS];
+
+  // 0 = ZERO_MV, MV
+  signed char mode_deltas[MAX_MODE_LF_DELTAS];
+  signed char last_mode_deltas[MAX_MODE_LF_DELTAS];
+};
+
+// Need to align this structure so when it is declared and
+// passed it can be loaded into vector registers.
+typedef struct {
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, mblim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, lim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, hev_thr[SIMD_WIDTH]);
+} loop_filter_thresh;
+
+typedef struct {
+  loop_filter_thresh lfthr[MAX_LOOP_FILTER + 1];
+  uint8_t lvl[MAX_SEGMENTS][MAX_REF_FRAMES][MAX_MODE_LF_DELTAS];
+} loop_filter_info_n;
+
+// This structure holds bit masks for all 8x8 blocks in a 64x64 region.
+// Each 1 bit represents a position in which we want to apply the loop filter.
+// Left_ entries refer to whether we apply a filter on the border to the
+// left of the block.   Above_ entries refer to whether or not to apply a
+// filter on the above border.   Int_ entries refer to whether or not to
+// apply borders on the 4x4 edges within the 8x8 block that each bit
+// represents.
+// Since each transform is accompanied by a potentially different type of
+// loop filter there is a different entry in the array for each transform size.
+typedef struct {
+  uint64_t left_y[TX_SIZES];
+  uint64_t above_y[TX_SIZES];
+  uint64_t int_4x4_y;
+  uint16_t left_uv[TX_SIZES];
+  uint16_t above_uv[TX_SIZES];
+  uint16_t int_4x4_uv;
+  uint8_t lfl_y[64];
+  uint8_t lfl_uv[16];
+} LOOP_FILTER_MASK;
+
+/* assorted loopfilter functions which get used elsewhere */
+struct VP9Common;
+struct macroblockd;
+struct VP9LfSyncData;
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+void vp9_setup_mask(struct VP9Common *const cm,
+                    const int mi_row, const int mi_col,
+                    MODE_INFO **mi_8x8, const int mode_info_stride,
+                    LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_ss00(struct VP9Common *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_ss11(struct VP9Common *const cm,
+                                 struct macroblockd_plane *const plane,
+                                 int mi_row,
+                                 LOOP_FILTER_MASK *lfm);
+
+void vp9_filter_block_plane_non420(struct VP9Common *cm,
+                                   struct macroblockd_plane *plane,
+                                   MODE_INFO **mi_8x8,
+                                   int mi_row, int mi_col);
+
+void vp9_loop_filter_init(struct VP9Common *cm);
+
+// Update the loop filter for the current frame.
+// This should be called before vp9_loop_filter_rows(), vp9_loop_filter_frame()
+// calls this function directly.
+void vp9_loop_filter_frame_init(struct VP9Common *cm, int default_filt_lvl);
+
+void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame,
+                           struct VP9Common *cm,
+                           struct macroblockd *mbd,
+                           int filter_level,
+                           int y_only, int partial_frame);
+
+// Apply the loop filter to [start, stop) macro block rows in frame_buffer.
+void vp9_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer,
+                          struct VP9Common *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
+                          int start, int stop, int y_only);
+
+typedef struct LoopFilterWorkerData {
+  YV12_BUFFER_CONFIG *frame_buffer;
+  struct VP9Common *cm;
+  struct macroblockd_plane planes[MAX_MB_PLANE];
+
+  int start;
+  int stop;
+  int y_only;
+} LFWorkerData;
+
+void vp9_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]);
+
+// Operates on the rows described by 'lf_data'.
+int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_LOOPFILTER_H_
diff --git a/media/libvpx/vp9/common/vp9_loopfilter_filters.c b/media/libvpx/vp9/common/vp9_loopfilter_filters.c
new file mode 100644
index 000000000..3cf4c3225
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_loopfilter_filters.c
@@ -0,0 +1,745 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+static INLINE int8_t signed_char_clamp(int t) {
+  return (int8_t)clamp(t, -128, 127);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE int16_t signed_char_clamp_high(int t, int bd) {
+  switch (bd) {
+    case 10:
+      return (int16_t)clamp(t, -128*4, 128*4-1);
+    case 12:
+      return (int16_t)clamp(t, -128*16, 128*16-1);
+    case 8:
+    default:
+      return (int16_t)clamp(t, -128, 128-1);
+  }
+}
+#endif
+
+// should we apply any filter at all: 11111111 yes, 00000000 no
+static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
+                                 uint8_t p3, uint8_t p2,
+                                 uint8_t p1, uint8_t p0,
+                                 uint8_t q0, uint8_t q1,
+                                 uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p3 - p2) > limit) * -1;
+  mask |= (abs(p2 - p1) > limit) * -1;
+  mask |= (abs(p1 - p0) > limit) * -1;
+  mask |= (abs(q1 - q0) > limit) * -1;
+  mask |= (abs(q2 - q1) > limit) * -1;
+  mask |= (abs(q3 - q2) > limit) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t flat_mask4(uint8_t thresh,
+                                uint8_t p3, uint8_t p2,
+                                uint8_t p1, uint8_t p0,
+                                uint8_t q0, uint8_t q1,
+                                uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p1 - p0) > thresh) * -1;
+  mask |= (abs(q1 - q0) > thresh) * -1;
+  mask |= (abs(p2 - p0) > thresh) * -1;
+  mask |= (abs(q2 - q0) > thresh) * -1;
+  mask |= (abs(p3 - p0) > thresh) * -1;
+  mask |= (abs(q3 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t flat_mask5(uint8_t thresh,
+                                uint8_t p4, uint8_t p3,
+                                uint8_t p2, uint8_t p1,
+                                uint8_t p0, uint8_t q0,
+                                uint8_t q1, uint8_t q2,
+                                uint8_t q3, uint8_t q4) {
+  int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
+  mask |= (abs(p4 - p0) > thresh) * -1;
+  mask |= (abs(q4 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+// is there high edge variance internal edge: 11111111 yes, 00000000 no
+static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
+                              uint8_t q0, uint8_t q1) {
+  int8_t hev = 0;
+  hev  |= (abs(p1 - p0) > thresh) * -1;
+  hev  |= (abs(q1 - q0) > thresh) * -1;
+  return hev;
+}
+
+static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
+                           uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
+  int8_t filter1, filter2;
+
+  const int8_t ps1 = (int8_t) *op1 ^ 0x80;
+  const int8_t ps0 = (int8_t) *op0 ^ 0x80;
+  const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
+  const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
+  const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
+
+  // add outer taps if we have high edge variance
+  int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
+
+  // inner taps
+  filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
+
+  // save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way
+  filter1 = signed_char_clamp(filter + 4) >> 3;
+  filter2 = signed_char_clamp(filter + 3) >> 3;
+
+  *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
+  *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
+
+  // outer tap adjustments
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
+  *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+}
+
+void vp9_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
+                            const uint8_t *blimit, const uint8_t *limit,
+                            const uint8_t *thresh, int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p],  q1 = s[1 * p],  q2 = s[2 * p],  q3 = s[3 * p];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
+    ++s;
+  }
+}
+
+void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  vp9_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1);
+  vp9_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh,
+                          int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0],  q1 = s[1],  q2 = s[2],  q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
+    s += pitch;
+  }
+}
+
+void vp9_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  vp9_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1);
+  vp9_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+                                  thresh1, 1);
+}
+
+static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                           uint8_t *op3, uint8_t *op2,
+                           uint8_t *op1, uint8_t *op0,
+                           uint8_t *oq0, uint8_t *oq1,
+                           uint8_t *oq2, uint8_t *oq3) {
+  if (flat && mask) {
+    const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    filter4(mask, thresh, op1,  op0, oq0, oq1);
+  }
+}
+
+void vp9_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
+                            const uint8_t *limit, const uint8_t *thresh,
+                            int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                                 s,         s + 1 * p, s + 2 * p, s + 3 * p);
+    ++s;
+  }
+}
+
+void vp9_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  vp9_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1);
+  vp9_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh,
+                          int count) {
+  int i;
+
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1,
+                                 s,     s + 1, s + 2, s + 3);
+    s += pitch;
+  }
+}
+
+void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  vp9_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1);
+  vp9_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+                                    thresh1, 1);
+}
+
+static INLINE void filter16(int8_t mask, uint8_t thresh,
+                            uint8_t flat, uint8_t flat2,
+                            uint8_t *op7, uint8_t *op6,
+                            uint8_t *op5, uint8_t *op4,
+                            uint8_t *op3, uint8_t *op2,
+                            uint8_t *op1, uint8_t *op0,
+                            uint8_t *oq0, uint8_t *oq1,
+                            uint8_t *oq2, uint8_t *oq3,
+                            uint8_t *oq4, uint8_t *oq5,
+                            uint8_t *oq6, uint8_t *oq7) {
+  if (flat2 && flat && mask) {
+    const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
+                  p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
+                  q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4, 4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5, 4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+    *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+    *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+  }
+}
+
+void vp9_lpf_horizontal_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh,
+                             int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat2 = flat_mask5(1,
+                             s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+                             q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
+
+    filter16(mask, *thresh, flat, flat2,
+             s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+             s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+             s,         s + 1 * p, s + 2 * p, s + 3 * p,
+             s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
+    ++s;
+  }
+}
+
+static void mb_lpf_vertical_edge_w(uint8_t *s, int p,
+                                   const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh,
+                                   int count) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1],  q2 = s[2], q3 = s[3];
+    const int8_t mask = filter_mask(*limit, *blimit,
+                                    p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                    q0, s[4], s[5], s[6], s[7]);
+
+    filter16(mask, *thresh, flat, flat2,
+             s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+             s,     s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
+    s += p;
+  }
+}
+
+void vp9_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                           const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
+}
+
+void vp9_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
+                                const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
+                                        uint16_t p3, uint16_t p2,
+                                        uint16_t p1, uint16_t p0,
+                                        uint16_t q0, uint16_t q1,
+                                        uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t limit16 = (uint16_t)limit << (bd - 8);
+  int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+  mask |= (abs(p3 - p2) > limit16) * -1;
+  mask |= (abs(p2 - p1) > limit16) * -1;
+  mask |= (abs(p1 - p0) > limit16) * -1;
+  mask |= (abs(q1 - q0) > limit16) * -1;
+  mask |= (abs(q2 - q1) > limit16) * -1;
+  mask |= (abs(q3 - q2) > limit16) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask4(uint8_t thresh,
+                                       uint16_t p3, uint16_t p2,
+                                       uint16_t p1, uint16_t p0,
+                                       uint16_t q0, uint16_t q1,
+                                       uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p1 - p0) > thresh16) * -1;
+  mask |= (abs(q1 - q0) > thresh16) * -1;
+  mask |= (abs(p2 - p0) > thresh16) * -1;
+  mask |= (abs(q2 - q0) > thresh16) * -1;
+  mask |= (abs(p3 - p0) > thresh16) * -1;
+  mask |= (abs(q3 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask5(uint8_t thresh,
+                                       uint16_t p4, uint16_t p3,
+                                       uint16_t p2, uint16_t p1,
+                                       uint16_t p0, uint16_t q0,
+                                       uint16_t q1, uint16_t q2,
+                                       uint16_t q3, uint16_t q4, int bd) {
+  int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p4 - p0) > thresh16) * -1;
+  mask |= (abs(q4 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+// Is there high edge variance internal edge:
+// 11111111_11111111 yes, 00000000_00000000 no ?
+static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
+                                      uint16_t q0, uint16_t q1, int bd) {
+  int16_t hev = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  hev |= (abs(p1 - p0) > thresh16) * -1;
+  hev |= (abs(q1 - q0) > thresh16) * -1;
+  return hev;
+}
+
+static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
+                                  uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+                                  int bd) {
+  int16_t filter1, filter2;
+  // ^0x80 equivalent to subtracting 0x80 from the values to turn them
+  // into -128 to +127 instead of 0 to 255.
+  int shift = bd - 8;
+  const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
+  const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
+  const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
+  const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
+  const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
+
+  // Add outer taps if we have high edge variance.
+  int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
+
+  // Inner taps.
+  filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
+
+  // Save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way.
+  filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
+  filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
+
+  *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
+  *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
+
+  // Outer tap adjustments.
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
+  *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
+}
+
+void vp9_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh, int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
+    ++s;
+  }
+}
+
+void vp9_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vp9_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0],  q1 = s[1],  q2 = s[2],  q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
+    s += pitch;
+  }
+}
+
+void vp9_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vp9_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, 1, bd);
+}
+
+static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                                  uint16_t *op3, uint16_t *op2,
+                                  uint16_t *op1, uint16_t *op0,
+                                  uint16_t *oq0, uint16_t *oq1,
+                                  uint16_t *oq2, uint16_t *oq3, int bd) {
+  if (flat && mask) {
+    const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    highbd_filter4(mask, thresh, op1,  op0, oq0, oq1, bd);
+  }
+}
+
+void vp9_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit, const uint8_t *thresh,
+                                   int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                         p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    highbd_filter8(mask, *thresh, flat,
+                 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                 s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
+    ++s;
+  }
+}
+
+void vp9_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p,
+                                        const uint8_t *blimit0,
+                                        const uint8_t *limit0,
+                                        const uint8_t *thresh0,
+                                        const uint8_t *blimit1,
+                                        const uint8_t *limit1,
+                                        const uint8_t *thresh1,
+                                        int bd) {
+  vp9_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int count, int bd) {
+  int i;
+
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    highbd_filter8(mask, *thresh, flat,
+                 s - 4, s - 3, s - 2, s - 1,
+                 s, s + 1, s + 2, s + 3,
+                 bd);
+    s += pitch;
+  }
+}
+
+void vp9_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch,
+                                      const uint8_t *blimit0,
+                                      const uint8_t *limit0,
+                                      const uint8_t *thresh0,
+                                      const uint8_t *blimit1,
+                                      const uint8_t *limit1,
+                                      const uint8_t *thresh1,
+                                      int bd) {
+  vp9_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+  vp9_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+                              thresh1, 1, bd);
+}
+
+static INLINE void highbd_filter16(int8_t mask, uint8_t thresh,
+                                   uint8_t flat, uint8_t flat2,
+                                   uint16_t *op7, uint16_t *op6,
+                                   uint16_t *op5, uint16_t *op4,
+                                   uint16_t *op3, uint16_t *op2,
+                                   uint16_t *op1, uint16_t *op0,
+                                   uint16_t *oq0, uint16_t *oq1,
+                                   uint16_t *oq2, uint16_t *oq3,
+                                   uint16_t *oq4, uint16_t *oq5,
+                                   uint16_t *oq6, uint16_t *oq7, int bd) {
+  if (flat2 && flat && mask) {
+    const uint16_t p7 = *op7;
+    const uint16_t p6 = *op6;
+    const uint16_t p5 = *op5;
+    const uint16_t p4 = *op4;
+    const uint16_t p3 = *op3;
+    const uint16_t p2 = *op2;
+    const uint16_t p1 = *op1;
+    const uint16_t p0 = *op0;
+    const uint16_t q0 = *oq0;
+    const uint16_t q1 = *oq1;
+    const uint16_t q2 = *oq2;
+    const uint16_t q3 = *oq3;
+    const uint16_t q4 = *oq4;
+    const uint16_t q5 = *oq5;
+    const uint16_t q6 = *oq6;
+    const uint16_t q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4, 4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5, 4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+    *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+    *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(p0 +
+                              q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
+                   bd);
+  }
+}
+
+void vp9_highbd_lpf_horizontal_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int count, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    const int8_t flat2 = highbd_flat_mask5(
+        1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+        q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2,
+                    s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+                    s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+                    s, s + 1 * p, s + 2 * p, s + 3 * p,
+                    s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p,
+                    bd);
+    ++s;
+  }
+}
+
+static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh,
+                                          int count, int bd) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint16_t p3 = s[-4];
+    const uint16_t p2 = s[-3];
+    const uint16_t p1 = s[-2];
+    const uint16_t p0 = s[-1];
+    const uint16_t q0 = s[0];
+    const uint16_t q1 = s[1];
+    const uint16_t q2 = s[2];
+    const uint16_t q3 = s[3];
+    const int8_t mask = highbd_filter_mask(*limit, *blimit,
+                                           p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+                                          bd);
+    const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                           q0, s[4], s[5], s[6], s[7], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2,
+                    s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+                    s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7,
+                    bd);
+    s += p;
+  }
+}
+
+void vp9_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh,
+                                  int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
+}
+
+void vp9_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
+                                       const uint8_t *blimit,
+                                       const uint8_t *limit,
+                                       const uint8_t *thresh,
+                                       int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/media/libvpx/vp9/common/vp9_mfqe.c b/media/libvpx/vp9/common/vp9_mfqe.c
new file mode 100644
index 000000000..bebb37eda
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_mfqe.c
@@ -0,0 +1,394 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_postproc.h"
+
+// TODO(jackychen): Replace this function with SSE2 code. There is
+// one SSE2 implementation in vp8, so will consider how to share it
+// between vp8 and vp9.
+static void filter_by_weight(const uint8_t *src, int src_stride,
+                             uint8_t *dst, int dst_stride,
+                             int block_size, int src_weight) {
+  const int dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  const int rounding_bit = 1 << (MFQE_PRECISION - 1);
+  int r, c;
+
+  for (r = 0; r < block_size; r++) {
+    for (c = 0; c < block_size; c++) {
+      dst[c] = (src[c] * src_weight + dst[c] * dst_weight + rounding_bit)
+               >> MFQE_PRECISION;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vp9_filter_by_weight8x8_c(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride, int src_weight) {
+  filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight);
+}
+
+void vp9_filter_by_weight16x16_c(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 int src_weight) {
+  filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight);
+}
+
+static void filter_by_weight32x32(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int weight) {
+  vp9_filter_by_weight16x16(src, src_stride, dst, dst_stride, weight);
+  vp9_filter_by_weight16x16(src + 16, src_stride, dst + 16, dst_stride,
+                            weight);
+  vp9_filter_by_weight16x16(src + src_stride * 16, src_stride,
+                            dst + dst_stride * 16, dst_stride, weight);
+  vp9_filter_by_weight16x16(src + src_stride * 16 + 16, src_stride,
+                            dst + dst_stride * 16 + 16, dst_stride, weight);
+}
+
+static void filter_by_weight64x64(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int weight) {
+  filter_by_weight32x32(src, src_stride, dst, dst_stride, weight);
+  filter_by_weight32x32(src + 32, src_stride, dst + 32,
+                        dst_stride, weight);
+  filter_by_weight32x32(src + src_stride * 32, src_stride,
+                        dst + dst_stride * 32, dst_stride, weight);
+  filter_by_weight32x32(src + src_stride * 32 + 32, src_stride,
+                        dst + dst_stride * 32 + 32, dst_stride, weight);
+}
+
+static void apply_ifactor(const uint8_t *y, int y_stride, uint8_t *yd,
+                          int yd_stride, const uint8_t *u, const uint8_t *v,
+                          int uv_stride, uint8_t *ud, uint8_t *vd,
+                          int uvd_stride, BLOCK_SIZE block_size,
+                          int weight) {
+  if (block_size == BLOCK_16X16) {
+    vp9_filter_by_weight16x16(y, y_stride, yd, yd_stride, weight);
+    vp9_filter_by_weight8x8(u, uv_stride, ud, uvd_stride, weight);
+    vp9_filter_by_weight8x8(v, uv_stride, vd, uvd_stride, weight);
+  } else if (block_size == BLOCK_32X32) {
+    filter_by_weight32x32(y, y_stride, yd, yd_stride, weight);
+    vp9_filter_by_weight16x16(u, uv_stride, ud, uvd_stride, weight);
+    vp9_filter_by_weight16x16(v, uv_stride, vd, uvd_stride, weight);
+  } else if (block_size == BLOCK_64X64) {
+    filter_by_weight64x64(y, y_stride, yd, yd_stride, weight);
+    filter_by_weight32x32(u, uv_stride, ud, uvd_stride, weight);
+    filter_by_weight32x32(v, uv_stride, vd, uvd_stride, weight);
+  }
+}
+
+// TODO(jackychen): Determine whether replace it with assembly code.
+static void copy_mem8x8(const uint8_t *src, int src_stride,
+                        uint8_t *dst, int dst_stride) {
+  int r;
+  for (r = 0; r < 8; r++) {
+    memcpy(dst, src, 8);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void copy_mem16x16(const uint8_t *src, int src_stride,
+                          uint8_t *dst, int dst_stride) {
+  int r;
+  for (r = 0; r < 16; r++) {
+    memcpy(dst, src, 16);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void copy_mem32x32(const uint8_t *src, int src_stride,
+                          uint8_t *dst, int dst_stride) {
+  copy_mem16x16(src, src_stride, dst, dst_stride);
+  copy_mem16x16(src + 16, src_stride, dst + 16, dst_stride);
+  copy_mem16x16(src + src_stride * 16, src_stride,
+                dst + dst_stride * 16, dst_stride);
+  copy_mem16x16(src + src_stride * 16 + 16, src_stride,
+                dst + dst_stride * 16 + 16, dst_stride);
+}
+
+void copy_mem64x64(const uint8_t *src, int src_stride,
+                   uint8_t *dst, int dst_stride) {
+  copy_mem32x32(src, src_stride, dst, dst_stride);
+  copy_mem32x32(src + 32, src_stride, dst + 32, dst_stride);
+  copy_mem32x32(src + src_stride * 32, src_stride,
+                dst + src_stride * 32, dst_stride);
+  copy_mem32x32(src + src_stride * 32 + 32, src_stride,
+                dst + src_stride * 32 + 32, dst_stride);
+}
+
+static void copy_block(const uint8_t *y, const uint8_t *u, const uint8_t *v,
+                       int y_stride, int uv_stride, uint8_t *yd, uint8_t *ud,
+                       uint8_t *vd, int yd_stride, int uvd_stride,
+                       BLOCK_SIZE bs) {
+  if (bs == BLOCK_16X16) {
+    copy_mem16x16(y, y_stride, yd, yd_stride);
+    copy_mem8x8(u, uv_stride, ud, uvd_stride);
+    copy_mem8x8(v, uv_stride, vd, uvd_stride);
+  } else if (bs == BLOCK_32X32) {
+    copy_mem32x32(y, y_stride, yd, yd_stride);
+    copy_mem16x16(u, uv_stride, ud, uvd_stride);
+    copy_mem16x16(v, uv_stride, vd, uvd_stride);
+  } else {
+    copy_mem64x64(y, y_stride, yd, yd_stride);
+    copy_mem32x32(u, uv_stride, ud, uvd_stride);
+    copy_mem32x32(v, uv_stride, vd, uvd_stride);
+  }
+}
+
+static void get_thr(BLOCK_SIZE bs, int qdiff, int *sad_thr, int *vdiff_thr) {
+  const int adj = qdiff >> MFQE_PRECISION;
+  if (bs == BLOCK_16X16) {
+    *sad_thr = 7 + adj;
+  } else if (bs == BLOCK_32X32) {
+    *sad_thr = 6 + adj;
+  } else {  // BLOCK_64X64
+    *sad_thr = 5 + adj;
+  }
+  *vdiff_thr = 125 + qdiff;
+}
+
+static void mfqe_block(BLOCK_SIZE bs, const uint8_t *y, const uint8_t *u,
+                       const uint8_t *v, int y_stride, int uv_stride,
+                       uint8_t *yd, uint8_t *ud, uint8_t *vd, int yd_stride,
+                       int uvd_stride, int qdiff) {
+  int sad, sad_thr, vdiff, vdiff_thr;
+  uint32_t sse;
+
+  get_thr(bs, qdiff, &sad_thr, &vdiff_thr);
+
+  if (bs == BLOCK_16X16) {
+    vdiff = (vpx_variance16x16(y, y_stride, yd, yd_stride, &sse) + 128) >> 8;
+    sad = (vpx_sad16x16(y, y_stride, yd, yd_stride) + 128) >> 8;
+  } else if (bs == BLOCK_32X32) {
+    vdiff = (vpx_variance32x32(y, y_stride, yd, yd_stride, &sse) + 512) >> 10;
+    sad = (vpx_sad32x32(y, y_stride, yd, yd_stride) + 512) >> 10;
+  } else /* if (bs == BLOCK_64X64) */ {
+    vdiff = (vpx_variance64x64(y, y_stride, yd, yd_stride, &sse) + 2048) >> 12;
+    sad = (vpx_sad64x64(y, y_stride, yd, yd_stride) + 2048) >> 12;
+  }
+
+  // vdiff > sad * 3 means vdiff should not be too small, otherwise,
+  // it might be a lighting change in smooth area. When there is a
+  // lighting change in smooth area, it is dangerous to do MFQE.
+  if (sad > 1 && vdiff > sad * 3) {
+    const int weight = 1 << MFQE_PRECISION;
+    int ifactor = weight * sad * vdiff / (sad_thr * vdiff_thr);
+    // When ifactor equals weight, no MFQE is done.
+    if (ifactor > weight) {
+      ifactor = weight;
+    }
+    apply_ifactor(y, y_stride, yd, yd_stride, u, v, uv_stride, ud, vd,
+                  uvd_stride, bs, ifactor);
+  } else {
+    // Copy the block from current frame (i.e., no mfqe is done).
+    copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
+               yd_stride, uvd_stride, bs);
+  }
+}
+
+static int mfqe_decision(MODE_INFO *mi, BLOCK_SIZE cur_bs) {
+  // Check the motion in current block(for inter frame),
+  // or check the motion in the correlated block in last frame (for keyframe).
+  const int mv_len_square = mi->mbmi.mv[0].as_mv.row *
+                            mi->mbmi.mv[0].as_mv.row +
+                            mi->mbmi.mv[0].as_mv.col *
+                            mi->mbmi.mv[0].as_mv.col;
+  const int mv_threshold = 100;
+  return mi->mbmi.mode >= NEARESTMV &&  // Not an intra block
+         cur_bs >= BLOCK_16X16 &&
+         mv_len_square <= mv_threshold;
+}
+
+// Process each partiton in a super block, recursively.
+static void mfqe_partition(VP9_COMMON *cm, MODE_INFO *mi, BLOCK_SIZE bs,
+                           const uint8_t *y, const uint8_t *u,
+                           const uint8_t *v, int y_stride, int uv_stride,
+                           uint8_t *yd, uint8_t *ud, uint8_t *vd,
+                           int yd_stride, int uvd_stride) {
+  int mi_offset, y_offset, uv_offset;
+  const BLOCK_SIZE cur_bs = mi->mbmi.sb_type;
+  const int qdiff = cm->base_qindex - cm->postproc_state.last_base_qindex;
+  const int bsl = b_width_log2_lookup[bs];
+  PARTITION_TYPE partition = partition_lookup[bsl][cur_bs];
+  const BLOCK_SIZE subsize = get_subsize(bs, partition);
+
+  if (cur_bs < BLOCK_8X8) {
+    // If there are blocks smaller than 8x8, it must be on the boundary.
+    return;
+  }
+  // No MFQE on blocks smaller than 16x16
+  if (bs == BLOCK_16X16) {
+    partition = PARTITION_NONE;
+  }
+  if (bs == BLOCK_64X64) {
+    mi_offset = 4;
+    y_offset = 32;
+    uv_offset = 16;
+  } else {
+    mi_offset = 2;
+    y_offset = 16;
+    uv_offset = 8;
+  }
+  switch (partition) {
+    BLOCK_SIZE mfqe_bs, bs_tmp;
+    case PARTITION_HORZ:
+      if (bs == BLOCK_64X64) {
+        mfqe_bs = BLOCK_64X32;
+        bs_tmp = BLOCK_32X32;
+      } else {
+        mfqe_bs = BLOCK_32X16;
+        bs_tmp = BLOCK_16X16;
+      }
+      if (mfqe_decision(mi, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
+                   y_stride, uv_stride, yd + y_offset, ud + uv_offset,
+                   vd + uv_offset, yd_stride, uvd_stride, qdiff);
+      }
+      if (mfqe_decision(mi + mi_offset * cm->mi_stride, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
+                   v + uv_offset * uv_stride, y_stride, uv_stride,
+                   yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                   vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
+                   u + uv_offset * uv_stride + uv_offset,
+                   v + uv_offset * uv_stride + uv_offset, y_stride,
+                   uv_stride, yd + y_offset * yd_stride + y_offset,
+                   ud + uv_offset * uvd_stride + uv_offset,
+                   vd + uv_offset * uvd_stride + uv_offset,
+                   yd_stride, uvd_stride, qdiff);
+      }
+      break;
+    case PARTITION_VERT:
+      if (bs == BLOCK_64X64) {
+        mfqe_bs = BLOCK_32X64;
+        bs_tmp = BLOCK_32X32;
+      } else {
+        mfqe_bs = BLOCK_16X32;
+        bs_tmp = BLOCK_16X16;
+      }
+      if (mfqe_decision(mi, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride, u + uv_offset * uv_stride,
+                   v + uv_offset * uv_stride, y_stride, uv_stride,
+                   yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                   vd + uv_offset * uvd_stride, yd_stride, uvd_stride, qdiff);
+      }
+      if (mfqe_decision(mi + mi_offset, mfqe_bs)) {
+        // Do mfqe on the first square partition.
+        mfqe_block(bs_tmp, y + y_offset, u + uv_offset, v + uv_offset,
+                   y_stride, uv_stride, yd + y_offset, ud + uv_offset,
+                   vd + uv_offset, yd_stride, uvd_stride, qdiff);
+        // Do mfqe on the second square partition.
+        mfqe_block(bs_tmp, y + y_offset * y_stride + y_offset,
+                   u + uv_offset * uv_stride + uv_offset,
+                   v + uv_offset * uv_stride + uv_offset, y_stride,
+                   uv_stride, yd + y_offset * yd_stride + y_offset,
+                   ud + uv_offset * uvd_stride + uv_offset,
+                   vd + uv_offset * uvd_stride + uv_offset,
+                   yd_stride, uvd_stride, qdiff);
+      }
+      break;
+    case PARTITION_NONE:
+      if (mfqe_decision(mi, cur_bs)) {
+        // Do mfqe on this partition.
+        mfqe_block(cur_bs, y, u, v, y_stride, uv_stride,
+                   yd, ud, vd, yd_stride, uvd_stride, qdiff);
+      } else {
+        // Copy the block from current frame(i.e., no mfqe is done).
+        copy_block(y, u, v, y_stride, uv_stride, yd, ud, vd,
+                   yd_stride, uvd_stride, bs);
+      }
+      break;
+    case PARTITION_SPLIT:
+      // Recursion on four square partitions, e.g. if bs is 64X64,
+      // then look into four 32X32 blocks in it.
+      mfqe_partition(cm, mi, subsize, y, u, v, y_stride, uv_stride, yd, ud, vd,
+                     yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset, subsize, y + y_offset, u + uv_offset,
+                     v + uv_offset, y_stride, uv_stride, yd + y_offset,
+                     ud + uv_offset, vd + uv_offset, yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset * cm->mi_stride, subsize,
+                     y + y_offset * y_stride, u + uv_offset * uv_stride,
+                     v + uv_offset * uv_stride, y_stride, uv_stride,
+                     yd + y_offset * yd_stride, ud + uv_offset * uvd_stride,
+                     vd + uv_offset * uvd_stride, yd_stride, uvd_stride);
+      mfqe_partition(cm, mi + mi_offset * cm->mi_stride + mi_offset,
+                     subsize, y + y_offset * y_stride + y_offset,
+                     u + uv_offset * uv_stride + uv_offset,
+                     v + uv_offset * uv_stride + uv_offset, y_stride,
+                     uv_stride, yd + y_offset * yd_stride + y_offset,
+                     ud + uv_offset * uvd_stride + uv_offset,
+                     vd + uv_offset * uvd_stride + uv_offset,
+                     yd_stride, uvd_stride);
+      break;
+    default:
+      assert(0);
+  }
+}
+
+void vp9_mfqe(VP9_COMMON *cm) {
+  int mi_row, mi_col;
+  // Current decoded frame.
+  const YV12_BUFFER_CONFIG *show = cm->frame_to_show;
+  // Last decoded frame and will store the MFQE result.
+  YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer;
+  // Loop through each super block.
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row += MI_BLOCK_SIZE) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
+      MODE_INFO *mi;
+      MODE_INFO *mi_local = cm->mi + (mi_row * cm->mi_stride + mi_col);
+      // Motion Info in last frame.
+      MODE_INFO *mi_prev = cm->postproc_state.prev_mi +
+                           (mi_row * cm->mi_stride + mi_col);
+      const uint32_t y_stride = show->y_stride;
+      const uint32_t uv_stride = show->uv_stride;
+      const uint32_t yd_stride = dest->y_stride;
+      const uint32_t uvd_stride = dest->uv_stride;
+      const uint32_t row_offset_y = mi_row << 3;
+      const uint32_t row_offset_uv = mi_row << 2;
+      const uint32_t col_offset_y = mi_col << 3;
+      const uint32_t col_offset_uv = mi_col << 2;
+      const uint8_t *y = show->y_buffer + row_offset_y * y_stride +
+                         col_offset_y;
+      const uint8_t *u = show->u_buffer + row_offset_uv * uv_stride +
+                         col_offset_uv;
+      const uint8_t *v = show->v_buffer + row_offset_uv * uv_stride +
+                         col_offset_uv;
+      uint8_t *yd = dest->y_buffer + row_offset_y * yd_stride + col_offset_y;
+      uint8_t *ud = dest->u_buffer + row_offset_uv * uvd_stride +
+                    col_offset_uv;
+      uint8_t *vd = dest->v_buffer + row_offset_uv * uvd_stride +
+                    col_offset_uv;
+      if (frame_is_intra_only(cm)) {
+        mi = mi_prev;
+      } else {
+        mi = mi_local;
+      }
+      mfqe_partition(cm, mi, BLOCK_64X64, y, u, v, y_stride, uv_stride, yd, ud,
+                     vd, yd_stride, uvd_stride);
+    }
+  }
+}
diff --git a/media/libvpx/vp9/common/vp9_mfqe.h b/media/libvpx/vp9/common/vp9_mfqe.h
new file mode 100644
index 000000000..dfff8c23d
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_mfqe.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_MFQE_H_
+#define VP9_COMMON_VP9_MFQE_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Multiframe Quality Enhancement.
+// The aim for MFQE is to replace pixel blocks in the current frame with
+// the correlated pixel blocks (with higher quality) in the last frame.
+// The replacement can only be taken in stationary blocks by checking
+// the motion of the blocks and other conditions such as the SAD of
+// the current block and correlated block, the variance of the block
+// difference, etc.
+void vp9_mfqe(struct VP9Common *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MFQE_H_
diff --git a/media/libvpx/vp9/common/vp9_mv.h b/media/libvpx/vp9/common/vp9_mv.h
new file mode 100644
index 000000000..5d89da8c2
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_mv.h
@@ -0,0 +1,55 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_MV_H_
+#define VP9_COMMON_VP9_MV_H_
+
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct mv {
+  int16_t row;
+  int16_t col;
+} MV;
+
+typedef union int_mv {
+  uint32_t as_int;
+  MV as_mv;
+} int_mv; /* facilitates faster equality tests and copies */
+
+typedef struct mv32 {
+  int32_t row;
+  int32_t col;
+} MV32;
+
+static INLINE int is_zero_mv(const MV *mv) {
+  return *((const uint32_t *)mv) == 0;
+}
+
+static INLINE int is_equal_mv(const MV *a, const MV *b) {
+  return  *((const uint32_t *)a) == *((const uint32_t *)b);
+}
+
+static INLINE void clamp_mv(MV *mv, int min_col, int max_col,
+                            int min_row, int max_row) {
+  mv->col = clamp(mv->col, min_col, max_col);
+  mv->row = clamp(mv->row, min_row, max_row);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MV_H_
diff --git a/media/libvpx/vp9/common/vp9_mvref_common.c b/media/libvpx/vp9/common/vp9_mvref_common.c
new file mode 100644
index 000000000..ce6952752
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_mvref_common.c
@@ -0,0 +1,228 @@
+
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_mvref_common.h"
+
+// This function searches the neighbourhood of a given MB/SB
+// to try and find candidate reference vectors.
+static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                             const TileInfo *const tile,
+                             MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                             int_mv *mv_ref_list,
+                             int block, int mi_row, int mi_col,
+                             find_mv_refs_sync sync, void *const data) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
+  int different_ref_found = 0;
+  int context_counter = 0;
+  const MV_REF *const  prev_frame_mvs = cm->use_prev_frame_mvs ?
+      cm->prev_frame->mvs + mi_row * cm->mi_cols + mi_col : NULL;
+
+  // Blank the reference vector list
+  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
+                                                   xd->mi_stride];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, block),
+                        refmv_count, mv_ref_list, Done);
+      else if (candidate->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 1, mv_ref->col, block),
+                        refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+      const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
+                                                    xd->mi_stride]->mbmi;
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[0], refmv_count, mv_ref_list, Done);
+      else if (candidate->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // TODO(hkuang): Remove this sync after fixing pthread_cond_broadcast
+  // on windows platform. The sync here is unncessary if use_perv_frame_mvs
+  // is 0. But after removing it, there will be hang in the unit test on windows
+  // due to several threads waiting for a thread's signal.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+    if (cm->frame_parallel_decode && sync != NULL) {
+      sync(data, mi_row);
+    }
+#endif
+
+  // Check the last frame's mode and mv info.
+  if (cm->use_prev_frame_mvs) {
+    // Synchronize here for frame parallel decode if sync function is provided.
+    if (cm->frame_parallel_decode && sync != NULL) {
+      sync(data, mi_row);
+    }
+
+    if (prev_frame_mvs->ref_frame[0] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[0], refmv_count, mv_ref_list, Done);
+    } else if (prev_frame_mvs->ref_frame[1] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[1], refmv_count, mv_ref_list, Done);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
+        const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
+                                              * xd->mi_stride]->mbmi;
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
+                                 refmv_count, mv_ref_list, Done);
+      }
+    }
+  }
+
+  // Since we still don't have a candidate we'll try the last frame.
+  if (cm->use_prev_frame_mvs) {
+    if (prev_frame_mvs->ref_frame[0] != ref_frame &&
+        prev_frame_mvs->ref_frame[0] > INTRA_FRAME) {
+      int_mv mv = prev_frame_mvs->mv[0];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[0]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done);
+    }
+
+    if (prev_frame_mvs->ref_frame[1] > INTRA_FRAME &&
+        prev_frame_mvs->ref_frame[1] != ref_frame &&
+        prev_frame_mvs->mv[1].as_int != prev_frame_mvs->mv[0].as_int) {
+      int_mv mv = prev_frame_mvs->mv[1];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[1]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done);
+    }
+  }
+
+ Done:
+
+  mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter];
+
+  // Clamp vectors
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
+    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
+}
+
+void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      const TileInfo *const tile,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                      int_mv *mv_ref_list,
+                      int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data) {
+  find_mv_refs_idx(cm, xd, tile, mi, ref_frame, mv_ref_list, -1,
+                   mi_row, mi_col, sync, data);
+}
+
+static void lower_mv_precision(MV *mv, int allow_hp) {
+  const int use_hp = allow_hp && vp9_use_mv_hp(mv);
+  if (!use_hp) {
+    if (mv->row & 1)
+      mv->row += (mv->row > 0 ? -1 : 1);
+    if (mv->col & 1)
+      mv->col += (mv->col > 0 ? -1 : 1);
+  }
+}
+
+void vp9_find_best_ref_mvs(MACROBLOCKD *xd, int allow_hp,
+                           int_mv *mvlist, int_mv *nearest_mv,
+                           int_mv *near_mv) {
+  int i;
+  // Make sure all the candidates are properly clamped etc
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i) {
+    lower_mv_precision(&mvlist[i].as_mv, allow_hp);
+    clamp_mv2(&mvlist[i].as_mv, xd);
+  }
+  *nearest_mv = mvlist[0];
+  *near_mv = mvlist[1];
+}
+
+void vp9_append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                   const TileInfo *const tile,
+                                   int block, int ref, int mi_row, int mi_col,
+                                   int_mv *nearest_mv, int_mv *near_mv) {
+  int_mv mv_list[MAX_MV_REF_CANDIDATES];
+  MODE_INFO *const mi = xd->mi[0];
+  b_mode_info *bmi = mi->bmi;
+  int n;
+
+  assert(MAX_MV_REF_CANDIDATES == 2);
+
+  find_mv_refs_idx(cm, xd, tile, mi, mi->mbmi.ref_frame[ref], mv_list, block,
+                   mi_row, mi_col, NULL, NULL);
+
+  near_mv->as_int = 0;
+  switch (block) {
+    case 0:
+      nearest_mv->as_int = mv_list[0].as_int;
+      near_mv->as_int = mv_list[1].as_int;
+      break;
+    case 1:
+    case 2:
+      nearest_mv->as_int = bmi[0].as_mv[ref].as_int;
+      for (n = 0; n < MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != mv_list[n].as_int) {
+          near_mv->as_int = mv_list[n].as_int;
+          break;
+        }
+      break;
+    case 3: {
+      int_mv candidates[2 + MAX_MV_REF_CANDIDATES];
+      candidates[0] = bmi[1].as_mv[ref];
+      candidates[1] = bmi[0].as_mv[ref];
+      candidates[2] = mv_list[0];
+      candidates[3] = mv_list[1];
+
+      nearest_mv->as_int = bmi[2].as_mv[ref].as_int;
+      for (n = 0; n < 2 + MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != candidates[n].as_int) {
+          near_mv->as_int = candidates[n].as_int;
+          break;
+        }
+      break;
+    }
+    default:
+      assert(0 && "Invalid block index.");
+  }
+}
diff --git a/media/libvpx/vp9/common/vp9_mvref_common.h b/media/libvpx/vp9/common/vp9_mvref_common.h
new file mode 100644
index 000000000..f1df52146
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_mvref_common.h
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP9_COMMON_VP9_MVREF_COMMON_H_
+#define VP9_COMMON_VP9_MVREF_COMMON_H_
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define LEFT_TOP_MARGIN ((VP9_ENC_BORDER_IN_PIXELS - VP9_INTERP_EXTEND) << 3)
+#define RIGHT_BOTTOM_MARGIN ((VP9_ENC_BORDER_IN_PIXELS -\
+                                VP9_INTERP_EXTEND) << 3)
+
+#define MVREF_NEIGHBOURS 8
+
+typedef struct position {
+  int row;
+  int col;
+} POSITION;
+
+typedef enum {
+  BOTH_ZERO = 0,
+  ZERO_PLUS_PREDICTED = 1,
+  BOTH_PREDICTED = 2,
+  NEW_PLUS_NON_INTRA = 3,
+  BOTH_NEW = 4,
+  INTRA_PLUS_NON_INTRA = 5,
+  BOTH_INTRA = 6,
+  INVALID_CASE = 9
+} motion_vector_context;
+
+// This is used to figure out a context for the ref blocks. The code flattens
+// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
+// adding 9 for each intra block, 3 for each zero mv and 1 for each new
+// motion vector. This single number is then converted into a context
+// with a single lookup ( counter_to_context ).
+static const int mode_2_counter[MB_MODE_COUNT] = {
+  9,  // DC_PRED
+  9,  // V_PRED
+  9,  // H_PRED
+  9,  // D45_PRED
+  9,  // D135_PRED
+  9,  // D117_PRED
+  9,  // D153_PRED
+  9,  // D207_PRED
+  9,  // D63_PRED
+  9,  // TM_PRED
+  0,  // NEARESTMV
+  0,  // NEARMV
+  3,  // ZEROMV
+  1,  // NEWMV
+};
+
+// There are 3^3 different combinations of 3 counts that can be either 0,1 or
+// 2. However the actual count can never be greater than 2 so the highest
+// counter we need is 18. 9 is an invalid counter that's never used.
+static const int counter_to_context[19] = {
+  BOTH_PREDICTED,  // 0
+  NEW_PLUS_NON_INTRA,  // 1
+  BOTH_NEW,  // 2
+  ZERO_PLUS_PREDICTED,  // 3
+  NEW_PLUS_NON_INTRA,  // 4
+  INVALID_CASE,  // 5
+  BOTH_ZERO,  // 6
+  INVALID_CASE,  // 7
+  INVALID_CASE,  // 8
+  INTRA_PLUS_NON_INTRA,  // 9
+  INTRA_PLUS_NON_INTRA,  // 10
+  INVALID_CASE,  // 11
+  INTRA_PLUS_NON_INTRA,  // 12
+  INVALID_CASE,  // 13
+  INVALID_CASE,  // 14
+  INVALID_CASE,  // 15
+  INVALID_CASE,  // 16
+  INVALID_CASE,  // 17
+  BOTH_INTRA  // 18
+};
+
+static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
+  // 4X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 4X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X4
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X8
+  {{-1, 0}, {0, -1}, {-1, -1}, {-2, 0}, {0, -2}, {-2, -1}, {-1, -2}, {-2, -2}},
+  // 8X16
+  {{0, -1}, {-1, 0}, {1, -1}, {-1, -1}, {0, -2}, {-2, 0}, {-2, -1}, {-1, -2}},
+  // 16X8
+  {{-1, 0}, {0, -1}, {-1, 1}, {-1, -1}, {-2, 0}, {0, -2}, {-1, -2}, {-2, -1}},
+  // 16X16
+  {{-1, 0}, {0, -1}, {-1, 1}, {1, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 16X32
+  {{0, -1}, {-1, 0}, {2, -1}, {-1, -1}, {-1, 1}, {0, -3}, {-3, 0}, {-3, -3}},
+  // 32X16
+  {{-1, 0}, {0, -1}, {-1, 2}, {-1, -1}, {1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X32
+  {{-1, 1}, {1, -1}, {-1, 2}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-3, -3}},
+  // 32X64
+  {{0, -1}, {-1, 0}, {4, -1}, {-1, 2}, {-1, -1}, {0, -3}, {-3, 0}, {2, -1}},
+  // 64X32
+  {{-1, 0}, {0, -1}, {-1, 4}, {2, -1}, {-1, -1}, {-3, 0}, {0, -3}, {-1, 2}},
+  // 64X64
+  {{-1, 3}, {3, -1}, {-1, 4}, {4, -1}, {-1, -1}, {-1, 0}, {0, -1}, {-1, 6}}
+};
+
+static const int idx_n_column_to_subblock[4][2] = {
+  {1, 2},
+  {1, 3},
+  {3, 2},
+  {3, 3}
+};
+
+// clamp_mv_ref
+#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
+
+static INLINE void clamp_mv_ref(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - MV_BORDER,
+               xd->mb_to_right_edge + MV_BORDER,
+               xd->mb_to_top_edge - MV_BORDER,
+               xd->mb_to_bottom_edge + MV_BORDER);
+}
+
+// This function returns either the appropriate sub block or block's mv
+// on whether the block_size < 8x8 and we have check_sub_blocks set.
+static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
+                                      int search_col, int block_idx) {
+  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
+          ? candidate->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
+              .as_mv[which_mv]
+          : candidate->mbmi.mv[which_mv];
+}
+
+
+// Performs mv sign inversion if indicated by the reference frame combination.
+static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
+                              const MV_REFERENCE_FRAME this_ref_frame,
+                              const int *ref_sign_bias) {
+  int_mv mv = mbmi->mv[ref];
+  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
+    mv.as_mv.row *= -1;
+    mv.as_mv.col *= -1;
+  }
+  return mv;
+}
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector it will also
+// skip all additional processing and jump to done!
+#define ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, Done) \
+  do { \
+    if (refmv_count) { \
+      if ((mv).as_int != (mv_ref_list)[0].as_int) { \
+        (mv_ref_list)[(refmv_count)] = (mv); \
+        goto Done; \
+      } \
+    } else { \
+      (mv_ref_list)[(refmv_count)++] = (mv); \
+    } \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV(mbmi, ref_frame, ref_sign_bias, refmv_count, \
+                                 mv_ref_list, Done) \
+  do { \
+    if (is_inter_block(mbmi)) { \
+      if ((mbmi)->ref_frame[0] != ref_frame) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias), \
+                        refmv_count, mv_ref_list, Done); \
+      if (has_second_ref(mbmi) && \
+          (mbmi)->ref_frame[1] != ref_frame && \
+          (mbmi)->mv[1].as_int != (mbmi)->mv[0].as_int) \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias), \
+                        refmv_count, mv_ref_list, Done); \
+    } \
+  } while (0)
+
+
+// Checks that the given mi_row, mi_col and search point
+// are inside the borders of the tile.
+static INLINE int is_inside(const TileInfo *const tile,
+                            int mi_col, int mi_row, int mi_rows,
+                            const POSITION *mi_pos) {
+  return !(mi_row + mi_pos->row < 0 ||
+           mi_col + mi_pos->col < tile->mi_col_start ||
+           mi_row + mi_pos->row >= mi_rows ||
+           mi_col + mi_pos->col >= tile->mi_col_end);
+}
+
+// TODO(jingning): this mv clamping function should be block size dependent.
+static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,
+               xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN,
+               xd->mb_to_top_edge - LEFT_TOP_MARGIN,
+               xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN);
+}
+
+typedef void (*find_mv_refs_sync)(void *const data, int mi_row);
+void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+                      const TileInfo *const tile,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                      int_mv *mv_ref_list, int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data);
+
+// check a list of motion vectors by sad score using a number rows of pixels
+// above and a number cols of pixels in the left to select the one with best
+// score to use as ref motion vector
+void vp9_find_best_ref_mvs(MACROBLOCKD *xd, int allow_hp,
+                           int_mv *mvlist, int_mv *nearest_mv, int_mv *near_mv);
+
+void vp9_append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                   const TileInfo *const tile,
+                                   int block, int ref, int mi_row, int mi_col,
+                                   int_mv *nearest_mv, int_mv *near_mv);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_MVREF_COMMON_H_
diff --git a/media/libvpx/vp9/common/vp9_onyxc_int.h b/media/libvpx/vp9/common/vp9_onyxc_int.h
new file mode 100644
index 000000000..3af2a41bd
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_onyxc_int.h
@@ -0,0 +1,452 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_ONYXC_INT_H_
+#define VP9_COMMON_VP9_ONYXC_INT_H_
+
+#include "./vpx_config.h"
+#include "vpx/internal/vpx_codec_internal.h"
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_frame_buffers.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_thread.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#if CONFIG_VP9_POSTPROC
+#include "vp9/common/vp9_postproc.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define REFS_PER_FRAME 3
+
+#define REF_FRAMES_LOG2 3
+#define REF_FRAMES (1 << REF_FRAMES_LOG2)
+
+// 4 scratch frames for the new frames to support a maximum of 4 cores decoding
+// in parallel, 3 for scaled references on the encoder.
+// TODO(hkuang): Add ondemand frame buffers instead of hardcoding the number
+// of framebuffers.
+// TODO(jkoleszar): These 3 extra references could probably come from the
+// normal reference pool.
+#define FRAME_BUFFERS (REF_FRAMES + 7)
+
+#define FRAME_CONTEXTS_LOG2 2
+#define FRAME_CONTEXTS (1 << FRAME_CONTEXTS_LOG2)
+
+#define NUM_PING_PONG_BUFFERS 2
+
+extern const struct {
+  PARTITION_CONTEXT above;
+  PARTITION_CONTEXT left;
+} partition_context_lookup[BLOCK_SIZES];
+
+
+typedef enum {
+  SINGLE_REFERENCE      = 0,
+  COMPOUND_REFERENCE    = 1,
+  REFERENCE_MODE_SELECT = 2,
+  REFERENCE_MODES       = 3,
+} REFERENCE_MODE;
+
+typedef struct {
+  int_mv mv[2];
+  MV_REFERENCE_FRAME ref_frame[2];
+} MV_REF;
+
+typedef struct {
+  int ref_count;
+  MV_REF *mvs;
+  int mi_rows;
+  int mi_cols;
+  vpx_codec_frame_buffer_t raw_frame_buffer;
+  YV12_BUFFER_CONFIG buf;
+
+  // The Following variables will only be used in frame parallel decode.
+
+  // frame_worker_owner indicates which FrameWorker owns this buffer. NULL means
+  // that no FrameWorker owns, or is decoding, this buffer.
+  VP9Worker *frame_worker_owner;
+
+  // row and col indicate which position frame has been decoded to in real
+  // pixel unit. They are reset to -1 when decoding begins and set to INT_MAX
+  // when the frame is fully decoded.
+  int row;
+  int col;
+} RefCntBuffer;
+
+typedef struct BufferPool {
+  // Protect BufferPool from being accessed by several FrameWorkers at
+  // the same time during frame parallel decode.
+  // TODO(hkuang): Try to use atomic variable instead of locking the whole pool.
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t pool_mutex;
+#endif
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  RefCntBuffer frame_bufs[FRAME_BUFFERS];
+
+  // Frame buffers allocated internally by the codec.
+  InternalFrameBufferList int_frame_buffers;
+} BufferPool;
+
+typedef struct VP9Common {
+  struct vpx_internal_error_info  error;
+  vpx_color_space_t color_space;
+  int width;
+  int height;
+  int display_width;
+  int display_height;
+  int last_width;
+  int last_height;
+
+  // TODO(jkoleszar): this implies chroma ss right now, but could vary per
+  // plane. Revisit as part of the future change to YV12_BUFFER_CONFIG to
+  // support additional planes.
+  int subsampling_x;
+  int subsampling_y;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  int use_highbitdepth;  // Marks if we need to use 16bit frame buffers.
+#endif
+
+  YV12_BUFFER_CONFIG *frame_to_show;
+  RefCntBuffer *prev_frame;
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd.
+  RefCntBuffer *cur_frame;
+
+  int ref_frame_map[REF_FRAMES]; /* maps fb_idx to reference slot */
+
+  // Prepare ref_frame_map for the next frame.
+  // Only used in frame parallel decode.
+  int next_ref_frame_map[REF_FRAMES];
+
+  // TODO(jkoleszar): could expand active_ref_idx to 4, with 0 as intra, and
+  // roll new_fb_idx into it.
+
+  // Each frame can reference REFS_PER_FRAME buffers
+  RefBuffer frame_refs[REFS_PER_FRAME];
+
+  int new_fb_idx;
+
+#if CONFIG_VP9_POSTPROC
+  YV12_BUFFER_CONFIG post_proc_buffer;
+  YV12_BUFFER_CONFIG post_proc_buffer_int;
+#endif
+
+  FRAME_TYPE last_frame_type;  /* last frame's frame type for motion search.*/
+  FRAME_TYPE frame_type;
+
+  int show_frame;
+  int last_show_frame;
+  int show_existing_frame;
+
+  // Flag signaling that the frame is encoded using only INTRA modes.
+  int intra_only;
+
+  int allow_high_precision_mv;
+
+  // Flag signaling that the frame context should be reset to default values.
+  // 0 or 1 implies don't reset, 2 reset just the context specified in the
+  // frame header, 3 reset all contexts.
+  int reset_frame_context;
+
+  // MBs, mb_rows/cols is in 16-pixel units; mi_rows/cols is in
+  // MODE_INFO (8-pixel) units.
+  int MBs;
+  int mb_rows, mi_rows;
+  int mb_cols, mi_cols;
+  int mi_stride;
+
+  /* profile settings */
+  TX_MODE tx_mode;
+
+  int base_qindex;
+  int y_dc_delta_q;
+  int uv_dc_delta_q;
+  int uv_ac_delta_q;
+  int16_t y_dequant[MAX_SEGMENTS][2];
+  int16_t uv_dequant[MAX_SEGMENTS][2];
+
+  /* We allocate a MODE_INFO struct for each macroblock, together with
+     an extra row on top and column on the left to simplify prediction. */
+  int mi_alloc_size;
+  MODE_INFO *mip; /* Base of allocated array */
+  MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
+
+  // TODO(agrange): Move prev_mi into encoder structure.
+  // prev_mip and prev_mi will only be allocated in VP9 encoder.
+  MODE_INFO *prev_mip; /* MODE_INFO array 'mip' from last decoded frame */
+  MODE_INFO *prev_mi;  /* 'mi' from last frame (points into prev_mip) */
+
+  // Separate mi functions between encoder and decoder.
+  int (*alloc_mi)(struct VP9Common *cm, int mi_size);
+  void (*free_mi)(struct VP9Common *cm);
+  void (*setup_mi)(struct VP9Common *cm);
+
+  // Grid of pointers to 8x8 MODE_INFO structs.  Any 8x8 not in the visible
+  // area will be NULL.
+  MODE_INFO **mi_grid_base;
+  MODE_INFO **mi_grid_visible;
+  MODE_INFO **prev_mi_grid_base;
+  MODE_INFO **prev_mi_grid_visible;
+
+  // Whether to use previous frame's motion vectors for prediction.
+  int use_prev_frame_mvs;
+
+  // Persistent mb segment id map used in prediction.
+  int seg_map_idx;
+  int prev_seg_map_idx;
+
+  uint8_t *seg_map_array[NUM_PING_PONG_BUFFERS];
+  uint8_t *last_frame_seg_map;
+  uint8_t *current_frame_seg_map;
+  int seg_map_alloc_size;
+
+  INTERP_FILTER interp_filter;
+
+  loop_filter_info_n lf_info;
+
+  int refresh_frame_context;    /* Two state 0 = NO, 1 = YES */
+
+  int ref_frame_sign_bias[MAX_REF_FRAMES];    /* Two state 0, 1 */
+
+  struct loopfilter lf;
+  struct segmentation seg;
+
+  // TODO(hkuang): Remove this as it is the same as frame_parallel_decode
+  // in pbi.
+  int frame_parallel_decode;  // frame-based threading.
+
+  // Context probabilities for reference frame prediction
+  MV_REFERENCE_FRAME comp_fixed_ref;
+  MV_REFERENCE_FRAME comp_var_ref[2];
+  REFERENCE_MODE reference_mode;
+
+  FRAME_CONTEXT *fc;  /* this frame entropy */
+  FRAME_CONTEXT *frame_contexts;   // FRAME_CONTEXTS
+  unsigned int  frame_context_idx; /* Context to use/update */
+  FRAME_COUNTS counts;
+
+  unsigned int current_video_frame;
+  BITSTREAM_PROFILE profile;
+
+  // VPX_BITS_8 in profile 0 or 1, VPX_BITS_10 or VPX_BITS_12 in profile 2 or 3.
+  vpx_bit_depth_t bit_depth;
+  vpx_bit_depth_t dequant_bit_depth;  // bit_depth of current dequantizer
+
+#if CONFIG_VP9_POSTPROC
+  struct postproc_state  postproc_state;
+#endif
+
+  int error_resilient_mode;
+  int frame_parallel_decoding_mode;
+
+  int log2_tile_cols, log2_tile_rows;
+  int byte_alignment;
+  int skip_loop_filter;
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  // Handles memory for the codec.
+  InternalFrameBufferList int_frame_buffers;
+
+  // External BufferPool passed from outside.
+  BufferPool *buffer_pool;
+
+  PARTITION_CONTEXT *above_seg_context;
+  ENTROPY_CONTEXT *above_context;
+  int above_context_alloc_cols;
+} VP9_COMMON;
+
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool);
+void unlock_buffer_pool(BufferPool *const pool);
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame(VP9_COMMON *cm, int index) {
+  if (index < 0 || index >= REF_FRAMES)
+    return NULL;
+  if (cm->ref_frame_map[index] < 0)
+    return NULL;
+  assert(cm->ref_frame_map[index] < FRAME_BUFFERS);
+  return &cm->buffer_pool->frame_bufs[cm->ref_frame_map[index]].buf;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(VP9_COMMON *cm) {
+  return &cm->buffer_pool->frame_bufs[cm->new_fb_idx].buf;
+}
+
+static INLINE int get_free_fb(VP9_COMMON *cm) {
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  int i;
+
+  lock_buffer_pool(cm->buffer_pool);
+  for (i = 0; i < FRAME_BUFFERS; ++i)
+    if (frame_bufs[i].ref_count == 0)
+      break;
+
+  if (i != FRAME_BUFFERS) {
+    frame_bufs[i].ref_count = 1;
+  } else {
+    // Reset i to be INVALID_IDX to indicate no free buffer found.
+    i = INVALID_IDX;
+  }
+
+  unlock_buffer_pool(cm->buffer_pool);
+  return i;
+}
+
+static INLINE void ref_cnt_fb(RefCntBuffer *bufs, int *idx, int new_idx) {
+  const int ref_index = *idx;
+
+  if (ref_index >= 0 && bufs[ref_index].ref_count > 0)
+    bufs[ref_index].ref_count--;
+
+  *idx = new_idx;
+
+  bufs[new_idx].ref_count++;
+}
+
+static INLINE int mi_cols_aligned_to_sb(int n_mis) {
+  return ALIGN_POWER_OF_TWO(n_mis, MI_BLOCK_SIZE_LOG2);
+}
+
+static INLINE void init_macroblockd(VP9_COMMON *cm, MACROBLOCKD *xd) {
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    xd->plane[i].dqcoeff = xd->dqcoeff;
+    xd->above_context[i] = cm->above_context +
+        i * sizeof(*cm->above_context) * 2 * mi_cols_aligned_to_sb(cm->mi_cols);
+
+    if (xd->plane[i].plane_type == PLANE_TYPE_Y) {
+      memcpy(xd->plane[i].seg_dequant, cm->y_dequant, sizeof(cm->y_dequant));
+    } else {
+      memcpy(xd->plane[i].seg_dequant, cm->uv_dequant, sizeof(cm->uv_dequant));
+    }
+    xd->fc = cm->fc;
+    xd->frame_parallel_decoding_mode = cm->frame_parallel_decoding_mode;
+  }
+
+  xd->above_seg_context = cm->above_seg_context;
+  xd->mi_stride = cm->mi_stride;
+  xd->error_info = &cm->error;
+}
+
+static INLINE int frame_is_intra_only(const VP9_COMMON *const cm) {
+  return cm->frame_type == KEY_FRAME || cm->intra_only;
+}
+
+static INLINE const vp9_prob* get_partition_probs(const VP9_COMMON *cm,
+                                                  int ctx) {
+  return frame_is_intra_only(cm) ? vp9_kf_partition_probs[ctx]
+                                 : cm->fc->partition_prob[ctx];
+}
+
+static INLINE void set_skip_context(MACROBLOCKD *xd, int mi_row, int mi_col) {
+  const int above_idx = mi_col * 2;
+  const int left_idx = (mi_row * 2) & 15;
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    pd->above_context = &xd->above_context[i][above_idx >> pd->subsampling_x];
+    pd->left_context = &xd->left_context[i][left_idx >> pd->subsampling_y];
+  }
+}
+
+static INLINE int calc_mi_size(int len) {
+  // len is in mi units.
+  return len + MI_BLOCK_SIZE;
+}
+
+static INLINE void set_mi_row_col(MACROBLOCKD *xd, const TileInfo *const tile,
+                                  int mi_row, int bh,
+                                  int mi_col, int bw,
+                                  int mi_rows, int mi_cols) {
+  xd->mb_to_top_edge    = -((mi_row * MI_SIZE) * 8);
+  xd->mb_to_bottom_edge = ((mi_rows - bh - mi_row) * MI_SIZE) * 8;
+  xd->mb_to_left_edge   = -((mi_col * MI_SIZE) * 8);
+  xd->mb_to_right_edge  = ((mi_cols - bw - mi_col) * MI_SIZE) * 8;
+
+  // Are edges available for intra prediction?
+  xd->up_available    = (mi_row != 0);
+  xd->left_available  = (mi_col > tile->mi_col_start);
+  if (xd->up_available) {
+    xd->above_mi = xd->mi[-xd->mi_stride];
+    // above_mi may be NULL in VP9 encoder's first pass.
+    xd->above_mbmi = xd->above_mi ? &xd->above_mi->mbmi : NULL;
+  } else {
+    xd->above_mi = NULL;
+    xd->above_mbmi = NULL;
+  }
+
+  if (xd->left_available) {
+    xd->left_mi = xd->mi[-1];
+    // left_mi may be NULL in VP9 encoder's first pass.
+    xd->left_mbmi = xd->left_mi ? &xd->left_mi->mbmi : NULL;
+  } else {
+    xd->left_mi = NULL;
+    xd->left_mbmi = NULL;
+  }
+}
+
+static INLINE void update_partition_context(MACROBLOCKD *xd,
+                                            int mi_row, int mi_col,
+                                            BLOCK_SIZE subsize,
+                                            BLOCK_SIZE bsize) {
+  PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
+  PARTITION_CONTEXT *const left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+
+  // num_4x4_blocks_wide_lookup[bsize] / 2
+  const int bs = num_8x8_blocks_wide_lookup[bsize];
+
+  // update the partition context at the end notes. set partition bits
+  // of block sizes larger than the current one to be one, and partition
+  // bits of smaller block sizes to be zero.
+  memset(above_ctx, partition_context_lookup[subsize].above, bs);
+  memset(left_ctx, partition_context_lookup[subsize].left, bs);
+}
+
+static INLINE int partition_plane_context(const MACROBLOCKD *xd,
+                                          int mi_row, int mi_col,
+                                          BLOCK_SIZE bsize) {
+  const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
+  const PARTITION_CONTEXT *left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
+  const int bsl = mi_width_log2_lookup[bsize];
+  int above = (*above_ctx >> bsl) & 1 , left = (*left_ctx >> bsl) & 1;
+
+  assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
+  assert(bsl >= 0);
+
+  return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_ONYXC_INT_H_
diff --git a/media/libvpx/vp9/common/vp9_postproc.c b/media/libvpx/vp9/common/vp9_postproc.c
new file mode 100644
index 000000000..d26a6eb5c
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_postproc.c
@@ -0,0 +1,745 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+#include "./vpx_config.h"
+#include "./vpx_scale_rtcd.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/mem.h"
+#include "vpx_scale/vpx_scale.h"
+#include "vpx_scale/yv12config.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_postproc.h"
+#include "vp9/common/vp9_systemdependent.h"
+#include "vp9/common/vp9_textblit.h"
+
+#if CONFIG_VP9_POSTPROC
+static const short kernel5[] = {
+  1, 1, 4, 1, 1
+};
+
+const short vp9_rv[] = {
+  8, 5, 2, 2, 8, 12, 4, 9, 8, 3,
+  0, 3, 9, 0, 0, 0, 8, 3, 14, 4,
+  10, 1, 11, 14, 1, 14, 9, 6, 12, 11,
+  8, 6, 10, 0, 0, 8, 9, 0, 3, 14,
+  8, 11, 13, 4, 2, 9, 0, 3, 9, 6,
+  1, 2, 3, 14, 13, 1, 8, 2, 9, 7,
+  3, 3, 1, 13, 13, 6, 6, 5, 2, 7,
+  11, 9, 11, 8, 7, 3, 2, 0, 13, 13,
+  14, 4, 12, 5, 12, 10, 8, 10, 13, 10,
+  4, 14, 4, 10, 0, 8, 11, 1, 13, 7,
+  7, 14, 6, 14, 13, 2, 13, 5, 4, 4,
+  0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+  8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+  3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+  3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+  13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+  5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+  9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+  4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+  3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+  11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+  5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+  0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+  10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+  4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+  0, 10, 0, 5, 13, 2, 12, 7, 11, 13,
+  8, 0, 4, 10, 7, 2, 7, 2, 2, 5,
+  3, 4, 7, 3, 3, 14, 14, 5, 9, 13,
+  3, 14, 3, 6, 3, 0, 11, 8, 13, 1,
+  13, 1, 12, 0, 10, 9, 7, 6, 2, 8,
+  5, 2, 13, 7, 1, 13, 14, 7, 6, 7,
+  9, 6, 10, 11, 7, 8, 7, 5, 14, 8,
+  4, 4, 0, 8, 7, 10, 0, 8, 14, 11,
+  3, 12, 5, 7, 14, 3, 14, 5, 2, 6,
+  11, 12, 12, 8, 0, 11, 13, 1, 2, 0,
+  5, 10, 14, 7, 8, 0, 4, 11, 0, 8,
+  0, 3, 10, 5, 8, 0, 11, 6, 7, 8,
+  10, 7, 13, 9, 2, 5, 1, 5, 10, 2,
+  4, 3, 5, 6, 10, 8, 9, 4, 11, 14,
+  3, 8, 3, 7, 8, 5, 11, 4, 12, 3,
+  11, 9, 14, 8, 14, 13, 4, 3, 1, 2,
+  14, 6, 5, 4, 4, 11, 4, 6, 2, 1,
+  5, 8, 8, 12, 13, 5, 14, 10, 12, 13,
+  0, 9, 5, 5, 11, 10, 13, 9, 10, 13,
+};
+
+static const uint8_t q_diff_thresh = 20;
+static const uint8_t last_q_thresh = 170;
+
+void vp9_post_proc_down_and_across_c(const uint8_t *src_ptr,
+                                     uint8_t *dst_ptr,
+                                     int src_pixels_per_line,
+                                     int dst_pixels_per_line,
+                                     int rows,
+                                     int cols,
+                                     int flimit) {
+  uint8_t const *p_src;
+  uint8_t *p_dst;
+  int row, col, i, v, kernel;
+  int pitch = src_pixels_per_line;
+  uint8_t d[8];
+  (void)dst_pixels_per_line;
+
+  for (row = 0; row < rows; row++) {
+    /* post_proc_down for one row */
+    p_src = src_ptr;
+    p_dst = dst_ptr;
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i * pitch]) > flimit)
+          goto down_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i * pitch];
+      }
+
+      v = (kernel >> 3);
+    down_skip_convolve:
+      p_dst[col] = v;
+    }
+
+    /* now post_proc_across */
+    p_src = dst_ptr;
+    p_dst = dst_ptr;
+
+    for (i = 0; i < 8; i++)
+      d[i] = p_src[i];
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      d[col & 7] = v;
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i]) > flimit)
+          goto across_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i];
+      }
+
+      d[col & 7] = (kernel >> 3);
+    across_skip_convolve:
+
+      if (col >= 2)
+        p_dst[col - 2] = d[(col - 2) & 7];
+    }
+
+    /* handle the last two pixels */
+    p_dst[col - 2] = d[(col - 2) & 7];
+    p_dst[col - 1] = d[(col - 1) & 7];
+
+
+    /* next row */
+    src_ptr += pitch;
+    dst_ptr += pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_post_proc_down_and_across_c(const uint16_t *src_ptr,
+                                            uint16_t *dst_ptr,
+                                            int src_pixels_per_line,
+                                            int dst_pixels_per_line,
+                                            int rows,
+                                            int cols,
+                                            int flimit) {
+  uint16_t const *p_src;
+  uint16_t *p_dst;
+  int row, col, i, v, kernel;
+  int pitch = src_pixels_per_line;
+  uint16_t d[8];
+
+  for (row = 0; row < rows; row++) {
+    // post_proc_down for one row.
+    p_src = src_ptr;
+    p_dst = dst_ptr;
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i * pitch]) > flimit)
+          goto down_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i * pitch];
+      }
+
+      v = (kernel >> 3);
+
+    down_skip_convolve:
+      p_dst[col] = v;
+    }
+
+    /* now post_proc_across */
+    p_src = dst_ptr;
+    p_dst = dst_ptr;
+
+    for (i = 0; i < 8; i++)
+      d[i] = p_src[i];
+
+    for (col = 0; col < cols; col++) {
+      kernel = 4;
+      v = p_src[col];
+
+      d[col & 7] = v;
+
+      for (i = -2; i <= 2; i++) {
+        if (abs(v - p_src[col + i]) > flimit)
+          goto across_skip_convolve;
+
+        kernel += kernel5[2 + i] * p_src[col + i];
+      }
+
+      d[col & 7] = (kernel >> 3);
+
+    across_skip_convolve:
+      if (col >= 2)
+        p_dst[col - 2] = d[(col - 2) & 7];
+    }
+
+    /* handle the last two pixels */
+    p_dst[col - 2] = d[(col - 2) & 7];
+    p_dst[col - 1] = d[(col - 1) & 7];
+
+
+    /* next row */
+    src_ptr += pitch;
+    dst_ptr += dst_pixels_per_line;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static int q2mbl(int x) {
+  if (x < 20) x = 20;
+
+  x = 50 + (x - 50) * 10 / 8;
+  return x * x / 3;
+}
+
+void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch,
+                                 int rows, int cols, int flimit) {
+  int r, c, i;
+  uint8_t *s = src;
+  uint8_t d[16];
+
+  for (r = 0; r < rows; r++) {
+    int sumsq = 0;
+    int sum = 0;
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i] * s[i];
+      sum += s[i];
+      d[i + 8] = 0;
+    }
+
+    for (c = 0; c < cols + 8; c++) {
+      int x = s[c + 7] - s[c - 8];
+      int y = s[c + 7] + s[c - 8];
+
+      sum += x;
+      sumsq += x * y;
+
+      d[c & 15] = s[c];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[c & 15] = (8 + sum + s[c]) >> 4;
+      }
+
+      s[c - 8] = d[(c - 8) & 15];
+    }
+    s += pitch;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_mbpost_proc_across_ip_c(uint16_t *src, int pitch,
+                                        int rows, int cols, int flimit) {
+  int r, c, i;
+
+  uint16_t *s = src;
+  uint16_t d[16];
+
+
+  for (r = 0; r < rows; r++) {
+    int sumsq = 0;
+    int sum   = 0;
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i] * s[i];
+      sum   += s[i];
+      d[i + 8] = 0;
+    }
+
+    for (c = 0; c < cols + 8; c++) {
+      int x = s[c + 7] - s[c - 8];
+      int y = s[c + 7] + s[c - 8];
+
+      sum  += x;
+      sumsq += x * y;
+
+      d[c & 15] = s[c];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[c & 15] = (8 + sum + s[c]) >> 4;
+      }
+
+      s[c - 8] = d[(c - 8) & 15];
+    }
+
+    s += pitch;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch,
+                            int rows, int cols, int flimit) {
+  int r, c, i;
+  const short *rv3 = &vp9_rv[63 & rand()]; // NOLINT
+
+  for (c = 0; c < cols; c++) {
+    uint8_t *s = &dst[c];
+    int sumsq = 0;
+    int sum   = 0;
+    uint8_t d[16];
+    const short *rv2 = rv3 + ((c * 17) & 127);
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i * pitch] * s[i * pitch];
+      sum   += s[i * pitch];
+    }
+
+    for (r = 0; r < rows + 8; r++) {
+      sumsq += s[7 * pitch] * s[ 7 * pitch] - s[-8 * pitch] * s[-8 * pitch];
+      sum  += s[7 * pitch] - s[-8 * pitch];
+      d[r & 15] = s[0];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[r & 15] = (rv2[r & 127] + sum + s[0]) >> 4;
+      }
+
+      s[-8 * pitch] = d[(r - 8) & 15];
+      s += pitch;
+    }
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_mbpost_proc_down_c(uint16_t *dst, int pitch,
+                                   int rows, int cols, int flimit) {
+  int r, c, i;
+  const int16_t *rv3 = &vp9_rv[63 & rand()];  // NOLINT
+
+  for (c = 0; c < cols; c++) {
+    uint16_t *s = &dst[c];
+    int sumsq = 0;
+    int sum = 0;
+    uint16_t d[16];
+    const int16_t *rv2 = rv3 + ((c * 17) & 127);
+
+    for (i = -8; i <= 6; i++) {
+      sumsq += s[i * pitch] * s[i * pitch];
+      sum += s[i * pitch];
+    }
+
+    for (r = 0; r < rows + 8; r++) {
+      sumsq += s[7 * pitch] * s[ 7 * pitch] - s[-8 * pitch] * s[-8 * pitch];
+      sum += s[7 * pitch] - s[-8 * pitch];
+      d[r & 15] = s[0];
+
+      if (sumsq * 15 - sum * sum < flimit) {
+        d[r & 15] = (rv2[r & 127] + sum + s[0]) >> 4;
+      }
+
+      s[-8 * pitch] = d[(r - 8) & 15];
+      s += pitch;
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void deblock_and_de_macro_block(YV12_BUFFER_CONFIG   *source,
+                                       YV12_BUFFER_CONFIG   *post,
+                                       int                   q,
+                                       int                   low_var_thresh,
+                                       int                   flag) {
+  double level = 6.0e-05 * q * q * q - .0067 * q * q + .306 * q + .0065;
+  int ppl = (int)(level + .5);
+  (void) low_var_thresh;
+  (void) flag;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (source->flags & YV12_FLAG_HIGHBITDEPTH) {
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->y_buffer),
+                                         CONVERT_TO_SHORTPTR(post->y_buffer),
+                                         source->y_stride, post->y_stride,
+                                         source->y_height, source->y_width,
+                                         ppl);
+
+    vp9_highbd_mbpost_proc_across_ip(CONVERT_TO_SHORTPTR(post->y_buffer),
+                                     post->y_stride, post->y_height,
+                                     post->y_width, q2mbl(q));
+
+    vp9_highbd_mbpost_proc_down(CONVERT_TO_SHORTPTR(post->y_buffer),
+                                post->y_stride, post->y_height,
+                                post->y_width, q2mbl(q));
+
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->u_buffer),
+                                         CONVERT_TO_SHORTPTR(post->u_buffer),
+                                         source->uv_stride, post->uv_stride,
+                                         source->uv_height, source->uv_width,
+                                         ppl);
+    vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(source->v_buffer),
+                                         CONVERT_TO_SHORTPTR(post->v_buffer),
+                                         source->uv_stride, post->uv_stride,
+                                         source->uv_height, source->uv_width,
+                                         ppl);
+  } else {
+    vp9_post_proc_down_and_across(source->y_buffer, post->y_buffer,
+                                  source->y_stride, post->y_stride,
+                                  source->y_height, source->y_width, ppl);
+
+    vp9_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
+                              post->y_width, q2mbl(q));
+
+    vp9_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
+                         post->y_width, q2mbl(q));
+
+    vp9_post_proc_down_and_across(source->u_buffer, post->u_buffer,
+                                  source->uv_stride, post->uv_stride,
+                                  source->uv_height, source->uv_width, ppl);
+    vp9_post_proc_down_and_across(source->v_buffer, post->v_buffer,
+                                  source->uv_stride, post->uv_stride,
+                                  source->uv_height, source->uv_width, ppl);
+  }
+#else
+  vp9_post_proc_down_and_across(source->y_buffer, post->y_buffer,
+                                source->y_stride, post->y_stride,
+                                source->y_height, source->y_width, ppl);
+
+  vp9_mbpost_proc_across_ip(post->y_buffer, post->y_stride, post->y_height,
+                            post->y_width, q2mbl(q));
+
+  vp9_mbpost_proc_down(post->y_buffer, post->y_stride, post->y_height,
+                       post->y_width, q2mbl(q));
+
+  vp9_post_proc_down_and_across(source->u_buffer, post->u_buffer,
+                                source->uv_stride, post->uv_stride,
+                                source->uv_height, source->uv_width, ppl);
+  vp9_post_proc_down_and_across(source->v_buffer, post->v_buffer,
+                                source->uv_stride, post->uv_stride,
+                                source->uv_height, source->uv_width, ppl);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                 int q) {
+  const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+                        + 0.0065 + 0.5);
+  int i;
+
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
+
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VP9_HIGHBITDEPTH
+    assert((src->flags & YV12_FLAG_HIGHBITDEPTH) ==
+           (dst->flags & YV12_FLAG_HIGHBITDEPTH));
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      vp9_highbd_post_proc_down_and_across(CONVERT_TO_SHORTPTR(srcs[i]),
+                                           CONVERT_TO_SHORTPTR(dsts[i]),
+                                           src_strides[i], dst_strides[i],
+                                           src_heights[i], src_widths[i], ppl);
+    } else {
+      vp9_post_proc_down_and_across(srcs[i], dsts[i],
+                                    src_strides[i], dst_strides[i],
+                                    src_heights[i], src_widths[i], ppl);
+    }
+#else
+    vp9_post_proc_down_and_across(srcs[i], dsts[i],
+                                  src_strides[i], dst_strides[i],
+                                  src_heights[i], src_widths[i], ppl);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+}
+
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                 int q) {
+  const int ppl = (int)(6.0e-05 * q * q * q - 0.0067 * q * q + 0.306 * q
+                        + 0.0065 + 0.5);
+  int i;
+
+  const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer};
+  const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride};
+  const int src_widths[3] = {src->y_width, src->uv_width, src->uv_width};
+  const int src_heights[3] = {src->y_height, src->uv_height, src->uv_height};
+
+  uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer};
+  const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride};
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    const int src_stride = src_strides[i];
+    const int src_width = src_widths[i] - 4;
+    const int src_height = src_heights[i] - 4;
+    const int dst_stride = dst_strides[i];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    assert((src->flags & YV12_FLAG_HIGHBITDEPTH) ==
+           (dst->flags & YV12_FLAG_HIGHBITDEPTH));
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      const uint16_t *const src_plane = CONVERT_TO_SHORTPTR(
+          srcs[i] + 2 * src_stride + 2);
+      uint16_t *const dst_plane = CONVERT_TO_SHORTPTR(
+          dsts[i] + 2 * dst_stride + 2);
+      vp9_highbd_post_proc_down_and_across(src_plane, dst_plane, src_stride,
+                                           dst_stride, src_height, src_width,
+                                           ppl);
+    } else {
+      const uint8_t *const src_plane = srcs[i] + 2 * src_stride + 2;
+      uint8_t *const dst_plane = dsts[i] + 2 * dst_stride + 2;
+
+      vp9_post_proc_down_and_across(src_plane, dst_plane, src_stride,
+                                    dst_stride, src_height, src_width, ppl);
+    }
+#else
+    const uint8_t *const src_plane = srcs[i] + 2 * src_stride + 2;
+    uint8_t *const dst_plane = dsts[i] + 2 * dst_stride + 2;
+    vp9_post_proc_down_and_across(src_plane, dst_plane, src_stride, dst_stride,
+                                  src_height, src_width, ppl);
+#endif
+  }
+}
+
+static double gaussian(double sigma, double mu, double x) {
+  return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+         (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+static void fillrd(struct postproc_state *state, int q, int a) {
+  char char_dist[300];
+
+  double sigma;
+  int ai = a, qi = q, i;
+
+  vp9_clear_system_state();
+
+  sigma = ai + .5 + .6 * (63 - qi) / 63.0;
+
+  /* set up a lookup table of 256 entries that matches
+   * a gaussian distribution with sigma determined by q.
+   */
+  {
+    int next, j;
+
+    next = 0;
+
+    for (i = -32; i < 32; i++) {
+      int a_i = (int)(0.5 + 256 * gaussian(sigma, 0, i));
+
+      if (a_i) {
+        for (j = 0; j < a_i; j++) {
+          char_dist[next + j] = (char) i;
+        }
+
+        next = next + j;
+      }
+    }
+
+    for (; next < 256; next++)
+      char_dist[next] = 0;
+  }
+
+  for (i = 0; i < 3072; i++) {
+    state->noise[i] = char_dist[rand() & 0xff];  // NOLINT
+  }
+
+  for (i = 0; i < 16; i++) {
+    state->blackclamp[i] = -char_dist[0];
+    state->whiteclamp[i] = -char_dist[0];
+    state->bothclamp[i] = -2 * char_dist[0];
+  }
+
+  state->last_q = q;
+  state->last_noise = a;
+}
+
+void vp9_plane_add_noise_c(uint8_t *start, char *noise,
+                           char blackclamp[16],
+                           char whiteclamp[16],
+                           char bothclamp[16],
+                           unsigned int width, unsigned int height, int pitch) {
+  unsigned int i, j;
+
+  // TODO(jbb): why does simd code use both but c doesn't,  normalize and
+  // fix..
+  (void) bothclamp;
+  for (i = 0; i < height; i++) {
+    uint8_t *pos = start + i * pitch;
+    char  *ref = (char *)(noise + (rand() & 0xff));  // NOLINT
+
+    for (j = 0; j < width; j++) {
+      if (pos[j] < blackclamp[0])
+        pos[j] = blackclamp[0];
+
+      if (pos[j] > 255 + whiteclamp[0])
+        pos[j] = 255 + whiteclamp[0];
+
+      pos[j] += ref[j];
+    }
+  }
+}
+
+static void swap_mi_and_prev_mi(VP9_COMMON *cm) {
+  // Current mip will be the prev_mip for the next frame.
+  MODE_INFO *temp = cm->postproc_state.prev_mip;
+  cm->postproc_state.prev_mip = cm->mip;
+  cm->mip = temp;
+
+  // Update the upper left visible macroblock ptrs.
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->postproc_state.prev_mi = cm->postproc_state.prev_mip + cm->mi_stride + 1;
+}
+
+int vp9_post_proc_frame(struct VP9Common *cm,
+                        YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *ppflags) {
+  const int q = MIN(105, cm->lf.filter_level * 2);
+  const int flags = ppflags->post_proc_flag;
+  YV12_BUFFER_CONFIG *const ppbuf = &cm->post_proc_buffer;
+  struct postproc_state *const ppstate = &cm->postproc_state;
+
+  if (!cm->frame_to_show)
+    return -1;
+
+  if (!flags) {
+    *dest = *cm->frame_to_show;
+    return 0;
+  }
+
+  vp9_clear_system_state();
+
+  // Alloc memory for prev_mip in the first frame.
+  if (cm->current_video_frame == 1) {
+    cm->postproc_state.last_base_qindex = cm->base_qindex;
+    cm->postproc_state.last_frame_valid = 1;
+    ppstate->prev_mip = vpx_calloc(cm->mi_alloc_size, sizeof(*cm->mip));
+    if (!ppstate->prev_mip) {
+      return 1;
+    }
+    ppstate->prev_mi = ppstate->prev_mip + cm->mi_stride + 1;
+    memset(ppstate->prev_mip, 0,
+           cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
+  }
+
+  // Allocate post_proc_buffer_int if needed.
+  if ((flags & VP9D_MFQE) && !cm->post_proc_buffer_int.buffer_alloc) {
+    if ((flags & VP9D_DEMACROBLOCK) || (flags & VP9D_DEBLOCK)) {
+      const int width = ALIGN_POWER_OF_TWO(cm->width, 4);
+      const int height = ALIGN_POWER_OF_TWO(cm->height, 4);
+
+      if (vp9_alloc_frame_buffer(&cm->post_proc_buffer_int, width, height,
+                                 cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cm->use_highbitdepth,
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 cm->byte_alignment) < 0) {
+        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate MFQE framebuffer");
+      }
+
+      // Ensure that postproc is set to all 0s so that post proc
+      // doesn't pull random data in from edge.
+      memset(cm->post_proc_buffer_int.buffer_alloc, 128,
+             cm->post_proc_buffer.frame_size);
+    }
+  }
+
+  if (vp9_realloc_frame_buffer(&cm->post_proc_buffer, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               VP9_DEC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL) < 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate post-processing buffer");
+
+  if ((flags & VP9D_MFQE) && cm->current_video_frame >= 2 &&
+      cm->postproc_state.last_frame_valid && cm->bit_depth == 8 &&
+      cm->postproc_state.last_base_qindex <= last_q_thresh &&
+      cm->base_qindex - cm->postproc_state.last_base_qindex >= q_diff_thresh) {
+    vp9_mfqe(cm);
+    // TODO(jackychen): Consider whether enable deblocking by default
+    // if mfqe is enabled. Need to take both the quality and the speed
+    // into consideration.
+    if ((flags & VP9D_DEMACROBLOCK) || (flags & VP9D_DEBLOCK)) {
+      vp8_yv12_copy_frame(ppbuf, &cm->post_proc_buffer_int);
+    }
+    if ((flags & VP9D_DEMACROBLOCK) && cm->post_proc_buffer_int.buffer_alloc) {
+      deblock_and_de_macro_block(&cm->post_proc_buffer_int, ppbuf,
+                                 q + (ppflags->deblocking_level - 5) * 10,
+                                 1, 0);
+    } else if (flags & VP9D_DEBLOCK) {
+      vp9_deblock(&cm->post_proc_buffer_int, ppbuf, q);
+    } else {
+      vp8_yv12_copy_frame(&cm->post_proc_buffer_int, ppbuf);
+    }
+  } else if (flags & VP9D_DEMACROBLOCK) {
+    deblock_and_de_macro_block(cm->frame_to_show, ppbuf,
+                               q + (ppflags->deblocking_level - 5) * 10, 1, 0);
+  } else if (flags & VP9D_DEBLOCK) {
+    vp9_deblock(cm->frame_to_show, ppbuf, q);
+  } else {
+    vp8_yv12_copy_frame(cm->frame_to_show, ppbuf);
+  }
+
+  cm->postproc_state.last_base_qindex = cm->base_qindex;
+  cm->postproc_state.last_frame_valid = 1;
+
+  if (flags & VP9D_ADDNOISE) {
+    const int noise_level = ppflags->noise_level;
+    if (ppstate->last_q != q ||
+        ppstate->last_noise != noise_level) {
+      fillrd(ppstate, 63 - q, noise_level);
+    }
+
+    vp9_plane_add_noise(ppbuf->y_buffer, ppstate->noise, ppstate->blackclamp,
+                        ppstate->whiteclamp, ppstate->bothclamp,
+                        ppbuf->y_width, ppbuf->y_height, ppbuf->y_stride);
+  }
+
+  *dest = *ppbuf;
+
+  /* handle problem with extending borders */
+  dest->y_width = cm->width;
+  dest->y_height = cm->height;
+  dest->uv_width = dest->y_width >> cm->subsampling_x;
+  dest->uv_height = dest->y_height >> cm->subsampling_y;
+
+  swap_mi_and_prev_mi(cm);
+  return 0;
+}
+#endif  // CONFIG_VP9_POSTPROC
diff --git a/media/libvpx/vp9/common/vp9_postproc.h b/media/libvpx/vp9/common/vp9_postproc.h
new file mode 100644
index 000000000..035c9cdf8
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_postproc.h
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_COMMON_VP9_POSTPROC_H_
+#define VP9_COMMON_VP9_POSTPROC_H_
+
+#include "vpx_ports/mem.h"
+#include "vpx_scale/yv12config.h"
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_mfqe.h"
+#include "vp9/common/vp9_ppflags.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct postproc_state {
+  int last_q;
+  int last_noise;
+  char noise[3072];
+  int last_base_qindex;
+  int last_frame_valid;
+  MODE_INFO *prev_mip;
+  MODE_INFO *prev_mi;
+  DECLARE_ALIGNED(16, char, blackclamp[16]);
+  DECLARE_ALIGNED(16, char, whiteclamp[16]);
+  DECLARE_ALIGNED(16, char, bothclamp[16]);
+};
+
+struct VP9Common;
+
+#define MFQE_PRECISION 4
+
+int vp9_post_proc_frame(struct VP9Common *cm,
+                        YV12_BUFFER_CONFIG *dest, vp9_ppflags_t *flags);
+
+void vp9_denoise(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
+
+void vp9_deblock(const YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst, int q);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_POSTPROC_H_
diff --git a/media/libvpx/vp9/common/vp9_ppflags.h b/media/libvpx/vp9/common/vp9_ppflags.h
new file mode 100644
index 000000000..12b989f43
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_ppflags.h
@@ -0,0 +1,43 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_PPFLAGS_H_
+#define VP9_COMMON_VP9_PPFLAGS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  VP9D_NOFILTERING            = 0,
+  VP9D_DEBLOCK                = 1 << 0,
+  VP9D_DEMACROBLOCK           = 1 << 1,
+  VP9D_ADDNOISE               = 1 << 2,
+  VP9D_DEBUG_TXT_FRAME_INFO   = 1 << 3,
+  VP9D_DEBUG_TXT_MBLK_MODES   = 1 << 4,
+  VP9D_DEBUG_TXT_DC_DIFF      = 1 << 5,
+  VP9D_DEBUG_TXT_RATE_INFO    = 1 << 6,
+  VP9D_DEBUG_DRAW_MV          = 1 << 7,
+  VP9D_DEBUG_CLR_BLK_MODES    = 1 << 8,
+  VP9D_DEBUG_CLR_FRM_REF_BLKS = 1 << 9,
+  VP9D_MFQE                   = 1 << 10
+};
+
+typedef struct {
+  int post_proc_flag;
+  int deblocking_level;
+  int noise_level;
+} vp9_ppflags_t;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_PPFLAGS_H_
diff --git a/media/libvpx/vp9/common/vp9_pred_common.c b/media/libvpx/vp9/common/vp9_pred_common.c
new file mode 100644
index 000000000..0aac4a9e6
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_pred_common.c
@@ -0,0 +1,381 @@
+
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <limits.h>
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+// Returns a context number for the given MB prediction signal
+int vp9_get_pred_context_switchable_interp(const MACROBLOCKD *xd) {
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int left_type = xd->left_available && is_inter_block(left_mbmi) ?
+                            left_mbmi->interp_filter : SWITCHABLE_FILTERS;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const int above_type = xd->up_available && is_inter_block(above_mbmi) ?
+                             above_mbmi->interp_filter : SWITCHABLE_FILTERS;
+
+  if (left_type == above_type)
+    return left_type;
+  else if (left_type == SWITCHABLE_FILTERS && above_type != SWITCHABLE_FILTERS)
+    return above_type;
+  else if (left_type != SWITCHABLE_FILTERS && above_type == SWITCHABLE_FILTERS)
+    return left_type;
+  else
+    return SWITCHABLE_FILTERS;
+}
+
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real macroblocks.
+// The prediction flags in these dummy entries are initialized to 0.
+// 0 - inter/inter, inter/--, --/inter, --/--
+// 1 - intra/inter, inter/intra
+// 2 - intra/--, --/intra
+// 3 - intra/intra
+int vp9_get_intra_inter_context(const MACROBLOCKD *xd) {
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+    return left_intra && above_intra ? 3
+                                     : left_intra || above_intra;
+  } else if (has_above || has_left) {  // one edge available
+    return 2 * !is_inter_block(has_above ? above_mbmi : left_mbmi);
+  } else {
+    return 0;
+  }
+}
+
+int vp9_get_reference_mode_context(const VP9_COMMON *cm,
+                                   const MACROBLOCKD *xd) {
+  int ctx;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    if (!has_second_ref(above_mbmi) && !has_second_ref(left_mbmi))
+      // neither edge uses comp pred (0/1)
+      ctx = (above_mbmi->ref_frame[0] == cm->comp_fixed_ref) ^
+            (left_mbmi->ref_frame[0] == cm->comp_fixed_ref);
+    else if (!has_second_ref(above_mbmi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (above_mbmi->ref_frame[0] == cm->comp_fixed_ref ||
+                 !is_inter_block(above_mbmi));
+    else if (!has_second_ref(left_mbmi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (left_mbmi->ref_frame[0] == cm->comp_fixed_ref ||
+                 !is_inter_block(left_mbmi));
+    else  // both edges use comp pred (4)
+      ctx = 4;
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!has_second_ref(edge_mbmi))
+      // edge does not use comp pred (0/1)
+      ctx = edge_mbmi->ref_frame[0] == cm->comp_fixed_ref;
+    else
+      // edge uses comp pred (3)
+      ctx = 3;
+  } else {  // no edges available (1)
+    ctx = 1;
+  }
+  assert(ctx >= 0 && ctx < COMP_INTER_CONTEXTS);
+  return ctx;
+}
+
+// Returns a context number for the given MB prediction signal
+int vp9_get_pred_context_comp_ref_p(const VP9_COMMON *cm,
+                                    const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  const int fix_ref_idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+  const int var_ref_idx = !fix_ref_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] != cm->comp_var_ref[1]);
+      else  // comp pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[var_ref_idx]
+                                    != cm->comp_var_ref[1]);
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mbmi);
+      const int a_sg = !has_second_ref(above_mbmi);
+      const MV_REFERENCE_FRAME vrfa = a_sg ? above_mbmi->ref_frame[0]
+                                           : above_mbmi->ref_frame[var_ref_idx];
+      const MV_REFERENCE_FRAME vrfl = l_sg ? left_mbmi->ref_frame[0]
+                                           : left_mbmi->ref_frame[var_ref_idx];
+
+      if (vrfa == vrfl && cm->comp_var_ref[1] == vrfa) {
+        pred_context = 0;
+      } else if (l_sg && a_sg) {  // single/single
+        if ((vrfa == cm->comp_fixed_ref && vrfl == cm->comp_var_ref[0]) ||
+            (vrfl == cm->comp_fixed_ref && vrfa == cm->comp_var_ref[0]))
+          pred_context = 4;
+        else if (vrfa == vrfl)
+          pred_context = 3;
+        else
+          pred_context = 1;
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME vrfc = l_sg ? vrfa : vrfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? vrfa : vrfl;
+        if (vrfc == cm->comp_var_ref[1] && rfs != cm->comp_var_ref[1])
+          pred_context = 1;
+        else if (rfs == cm->comp_var_ref[1] && vrfc != cm->comp_var_ref[1])
+          pred_context = 2;
+        else
+          pred_context = 4;
+      } else if (vrfa == vrfl) {  // comp/comp
+        pred_context = 4;
+      } else {
+        pred_context = 2;
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi))
+        pred_context = 4 * (edge_mbmi->ref_frame[var_ref_idx]
+                              != cm->comp_var_ref[1]);
+      else
+        pred_context = 3 * (edge_mbmi->ref_frame[0] != cm->comp_var_ref[1]);
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+int vp9_get_pred_context_single_ref_p1(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi))
+        pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST_FRAME);
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        pred_context = 1 + (above0 == LAST_FRAME || above1 == LAST_FRAME ||
+                            left0 == LAST_FRAME || left1 == LAST_FRAME);
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == LAST_FRAME)
+          pred_context = 3 + (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+        else
+          pred_context = (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+      } else {
+        pred_context = 2 * (above0 == LAST_FRAME) + 2 * (left0 == LAST_FRAME);
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+    if (!is_inter_block(edge_mbmi)) {  // intra
+      pred_context = 2;
+    } else {  // inter
+      if (!has_second_ref(edge_mbmi))
+        pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST_FRAME);
+    }
+  } else {  // no edges available
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+int vp9_get_pred_context_single_ref_p2(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {
+        if (edge_mbmi->ref_frame[0] == LAST_FRAME)
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME);
+      } else {
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME ||
+                                edge_mbmi->ref_frame[1] == GOLDEN_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        if (above0 == left0 && above1 == left1)
+          pred_context = 3 * (above0 == GOLDEN_FRAME ||
+                              above1 == GOLDEN_FRAME ||
+                              left0 == GOLDEN_FRAME ||
+                              left1 == GOLDEN_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == GOLDEN_FRAME)
+          pred_context = 3 + (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+        else if (rfs == ALTREF_FRAME)
+          pred_context = crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME;
+        else
+          pred_context = 1 + 2 * (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+      } else {
+        if (above0 == LAST_FRAME && left0 == LAST_FRAME) {
+          pred_context = 3;
+        } else if (above0 == LAST_FRAME || left0 == LAST_FRAME) {
+          const MV_REFERENCE_FRAME edge0 = (above0 == LAST_FRAME) ? left0
+                                                                  : above0;
+          pred_context = 4 * (edge0 == GOLDEN_FRAME);
+        } else {
+          pred_context = 2 * (above0 == GOLDEN_FRAME) +
+                             2 * (left0 == GOLDEN_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (edge_mbmi->ref_frame[0] == LAST_FRAME && !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))
+      pred_context = 4 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME);
+    else
+      pred_context = 3 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME ||
+                          edge_mbmi->ref_frame[1] == GOLDEN_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+// Returns a context number for the given MB prediction signal
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real blocks.
+// The prediction flags in these dummy entries are initialized to 0.
+int vp9_get_tx_size_context(const MACROBLOCKD *xd) {
+  const int max_tx_size = max_txsize_lookup[xd->mi[0]->mbmi.sb_type];
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+  int above_ctx = (has_above && !above_mbmi->skip) ? (int)above_mbmi->tx_size
+                                                   : max_tx_size;
+  int left_ctx = (has_left && !left_mbmi->skip) ? (int)left_mbmi->tx_size
+                                                : max_tx_size;
+  if (!has_left)
+    left_ctx = above_ctx;
+
+  if (!has_above)
+    above_ctx = left_ctx;
+
+  return (above_ctx + left_ctx) > max_tx_size;
+}
+
+int vp9_get_segment_id(const VP9_COMMON *cm, const uint8_t *segment_ids,
+                       BLOCK_SIZE bsize, int mi_row, int mi_col) {
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = MIN(cm->mi_cols - mi_col, bw);
+  const int ymis = MIN(cm->mi_rows - mi_row, bh);
+  int x, y, segment_id = INT_MAX;
+
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++)
+      segment_id = MIN(segment_id,
+                       segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
diff --git a/media/libvpx/vp9/common/vp9_pred_common.h b/media/libvpx/vp9/common/vp9_pred_common.h
new file mode 100644
index 000000000..bc19d28b9
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_pred_common.h
@@ -0,0 +1,133 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_PRED_COMMON_H_
+#define VP9_COMMON_VP9_PRED_COMMON_H_
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int vp9_get_segment_id(const VP9_COMMON *cm, const uint8_t *segment_ids,
+                       BLOCK_SIZE bsize, int mi_row, int mi_col);
+
+static INLINE int vp9_get_pred_context_seg_id(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_sip = (above_mi != NULL) ?
+                        above_mi->mbmi.seg_id_predicted : 0;
+  const int left_sip = (left_mi != NULL) ? left_mi->mbmi.seg_id_predicted : 0;
+
+  return above_sip + left_sip;
+}
+
+static INLINE vp9_prob vp9_get_pred_prob_seg_id(const struct segmentation *seg,
+                                                const MACROBLOCKD *xd) {
+  return seg->pred_probs[vp9_get_pred_context_seg_id(xd)];
+}
+
+static INLINE int vp9_get_skip_context(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_skip = (above_mi != NULL) ? above_mi->mbmi.skip : 0;
+  const int left_skip = (left_mi != NULL) ? left_mi->mbmi.skip : 0;
+  return above_skip + left_skip;
+}
+
+static INLINE vp9_prob vp9_get_skip_prob(const VP9_COMMON *cm,
+                                         const MACROBLOCKD *xd) {
+  return cm->fc->skip_probs[vp9_get_skip_context(xd)];
+}
+
+int vp9_get_pred_context_switchable_interp(const MACROBLOCKD *xd);
+
+int vp9_get_intra_inter_context(const MACROBLOCKD *xd);
+
+static INLINE vp9_prob vp9_get_intra_inter_prob(const VP9_COMMON *cm,
+                                                const MACROBLOCKD *xd) {
+  return cm->fc->intra_inter_prob[vp9_get_intra_inter_context(xd)];
+}
+
+int vp9_get_reference_mode_context(const VP9_COMMON *cm, const MACROBLOCKD *xd);
+
+static INLINE vp9_prob vp9_get_reference_mode_prob(const VP9_COMMON *cm,
+                                                   const MACROBLOCKD *xd) {
+  return cm->fc->comp_inter_prob[vp9_get_reference_mode_context(cm, xd)];
+}
+
+int vp9_get_pred_context_comp_ref_p(const VP9_COMMON *cm,
+                                    const MACROBLOCKD *xd);
+
+static INLINE vp9_prob vp9_get_pred_prob_comp_ref_p(const VP9_COMMON *cm,
+                                                    const MACROBLOCKD *xd) {
+  const int pred_context = vp9_get_pred_context_comp_ref_p(cm, xd);
+  return cm->fc->comp_ref_prob[pred_context];
+}
+
+int vp9_get_pred_context_single_ref_p1(const MACROBLOCKD *xd);
+
+static INLINE vp9_prob vp9_get_pred_prob_single_ref_p1(const VP9_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[vp9_get_pred_context_single_ref_p1(xd)][0];
+}
+
+int vp9_get_pred_context_single_ref_p2(const MACROBLOCKD *xd);
+
+static INLINE vp9_prob vp9_get_pred_prob_single_ref_p2(const VP9_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[vp9_get_pred_context_single_ref_p2(xd)][1];
+}
+
+int vp9_get_tx_size_context(const MACROBLOCKD *xd);
+
+static INLINE const vp9_prob *get_tx_probs(TX_SIZE max_tx_size, int ctx,
+                                           const struct tx_probs *tx_probs) {
+  switch (max_tx_size) {
+    case TX_8X8:
+      return tx_probs->p8x8[ctx];
+    case TX_16X16:
+      return tx_probs->p16x16[ctx];
+    case TX_32X32:
+      return tx_probs->p32x32[ctx];
+    default:
+      assert(0 && "Invalid max_tx_size.");
+      return NULL;
+  }
+}
+
+static INLINE const vp9_prob *get_tx_probs2(TX_SIZE max_tx_size,
+                                            const MACROBLOCKD *xd,
+                                            const struct tx_probs *tx_probs) {
+  return get_tx_probs(max_tx_size, vp9_get_tx_size_context(xd), tx_probs);
+}
+
+static INLINE unsigned int *get_tx_counts(TX_SIZE max_tx_size, int ctx,
+                                          struct tx_counts *tx_counts) {
+  switch (max_tx_size) {
+    case TX_8X8:
+      return tx_counts->p8x8[ctx];
+    case TX_16X16:
+      return tx_counts->p16x16[ctx];
+    case TX_32X32:
+      return tx_counts->p32x32[ctx];
+    default:
+      assert(0 && "Invalid max_tx_size.");
+      return NULL;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_PRED_COMMON_H_
diff --git a/media/libvpx/vp9/common/vp9_prob.c b/media/libvpx/vp9/common/vp9_prob.c
new file mode 100644
index 000000000..3b7b9bf3b
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_prob.c
@@ -0,0 +1,53 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_prob.h"
+
+const uint8_t vp9_norm[256] = {
+  0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4,
+  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+};
+
+static unsigned int tree_merge_probs_impl(unsigned int i,
+                                          const vp9_tree_index *tree,
+                                          const vp9_prob *pre_probs,
+                                          const unsigned int *counts,
+                                          vp9_prob *probs) {
+  const int l = tree[i];
+  const unsigned int left_count = (l <= 0)
+                 ? counts[-l]
+                 : tree_merge_probs_impl(l, tree, pre_probs, counts, probs);
+  const int r = tree[i + 1];
+  const unsigned int right_count = (r <= 0)
+                 ? counts[-r]
+                 : tree_merge_probs_impl(r, tree, pre_probs, counts, probs);
+  const unsigned int ct[2] = { left_count, right_count };
+  probs[i >> 1] = mode_mv_merge_probs(pre_probs[i >> 1], ct);
+  return left_count + right_count;
+}
+
+void vp9_tree_merge_probs(const vp9_tree_index *tree, const vp9_prob *pre_probs,
+                          const unsigned int *counts, vp9_prob *probs) {
+  tree_merge_probs_impl(0, tree, pre_probs, counts, probs);
+}
diff --git a/media/libvpx/vp9/common/vp9_prob.h b/media/libvpx/vp9/common/vp9_prob.h
new file mode 100644
index 000000000..c69c62c81
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_prob.h
@@ -0,0 +1,104 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_PROB_H_
+#define VP9_COMMON_VP9_PROB_H_
+
+#include "./vpx_config.h"
+
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef uint8_t vp9_prob;
+
+#define MAX_PROB 255
+
+#define vp9_prob_half ((vp9_prob) 128)
+
+typedef int8_t vp9_tree_index;
+
+#define TREE_SIZE(leaf_count) (2 * (leaf_count) - 2)
+
+#define vp9_complement(x) (255 - x)
+
+#define MODE_MV_COUNT_SAT 20
+
+/* We build coding trees compactly in arrays.
+   Each node of the tree is a pair of vp9_tree_indices.
+   Array index often references a corresponding probability table.
+   Index <= 0 means done encoding/decoding and value = -Index,
+   Index > 0 means need another bit, specification at index.
+   Nonnegative indices are always even;  processing begins at node 0. */
+
+typedef const vp9_tree_index vp9_tree[];
+
+static INLINE vp9_prob clip_prob(int p) {
+  return (p > 255) ? 255 : (p < 1) ? 1 : p;
+}
+
+static INLINE vp9_prob get_prob(int num, int den) {
+  return (den == 0) ? 128u : clip_prob(((int64_t)num * 256 + (den >> 1)) / den);
+}
+
+static INLINE vp9_prob get_binary_prob(int n0, int n1) {
+  return get_prob(n0, n0 + n1);
+}
+
+/* This function assumes prob1 and prob2 are already within [1,255] range. */
+static INLINE vp9_prob weighted_prob(int prob1, int prob2, int factor) {
+  return ROUND_POWER_OF_TWO(prob1 * (256 - factor) + prob2 * factor, 8);
+}
+
+static INLINE vp9_prob merge_probs(vp9_prob pre_prob,
+                                   const unsigned int ct[2],
+                                   unsigned int count_sat,
+                                   unsigned int max_update_factor) {
+  const vp9_prob prob = get_binary_prob(ct[0], ct[1]);
+  const unsigned int count = MIN(ct[0] + ct[1], count_sat);
+  const unsigned int factor = max_update_factor * count / count_sat;
+  return weighted_prob(pre_prob, prob, factor);
+}
+
+// MODE_MV_MAX_UPDATE_FACTOR (128) * count / MODE_MV_COUNT_SAT;
+static const int count_to_update_factor[MODE_MV_COUNT_SAT + 1] = {
+  0, 6, 12, 19, 25, 32, 38, 44, 51, 57, 64,
+  70, 76, 83, 89, 96, 102, 108, 115, 121, 128
+};
+
+static INLINE vp9_prob mode_mv_merge_probs(vp9_prob pre_prob,
+                                           const unsigned int ct[2]) {
+  const unsigned int den = ct[0] + ct[1];
+  if (den == 0) {
+    return pre_prob;
+  } else {
+    const unsigned int count = MIN(den, MODE_MV_COUNT_SAT);
+    const unsigned int factor = count_to_update_factor[count];
+    const vp9_prob prob =
+        clip_prob(((int64_t)(ct[0]) * 256 + (den >> 1)) / den);
+    return weighted_prob(pre_prob, prob, factor);
+  }
+}
+
+void vp9_tree_merge_probs(const vp9_tree_index *tree, const vp9_prob *pre_probs,
+                          const unsigned int *counts, vp9_prob *probs);
+
+
+DECLARE_ALIGNED(16, extern const uint8_t, vp9_norm[256]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_PROB_H_
diff --git a/media/libvpx/vp9/common/vp9_quant_common.c b/media/libvpx/vp9/common/vp9_quant_common.c
new file mode 100644
index 000000000..564a3eb0c
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_quant_common.c
@@ -0,0 +1,278 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_seg_common.h"
+
+static const int16_t dc_qlookup[QINDEX_RANGE] = {
+  4,       8,    8,    9,   10,   11,   12,   12,
+  13,     14,   15,   16,   17,   18,   19,   19,
+  20,     21,   22,   23,   24,   25,   26,   26,
+  27,     28,   29,   30,   31,   32,   32,   33,
+  34,     35,   36,   37,   38,   38,   39,   40,
+  41,     42,   43,   43,   44,   45,   46,   47,
+  48,     48,   49,   50,   51,   52,   53,   53,
+  54,     55,   56,   57,   57,   58,   59,   60,
+  61,     62,   62,   63,   64,   65,   66,   66,
+  67,     68,   69,   70,   70,   71,   72,   73,
+  74,     74,   75,   76,   77,   78,   78,   79,
+  80,     81,   81,   82,   83,   84,   85,   85,
+  87,     88,   90,   92,   93,   95,   96,   98,
+  99,    101,  102,  104,  105,  107,  108,  110,
+  111,   113,  114,  116,  117,  118,  120,  121,
+  123,   125,  127,  129,  131,  134,  136,  138,
+  140,   142,  144,  146,  148,  150,  152,  154,
+  156,   158,  161,  164,  166,  169,  172,  174,
+  177,   180,  182,  185,  187,  190,  192,  195,
+  199,   202,  205,  208,  211,  214,  217,  220,
+  223,   226,  230,  233,  237,  240,  243,  247,
+  250,   253,  257,  261,  265,  269,  272,  276,
+  280,   284,  288,  292,  296,  300,  304,  309,
+  313,   317,  322,  326,  330,  335,  340,  344,
+  349,   354,  359,  364,  369,  374,  379,  384,
+  389,   395,  400,  406,  411,  417,  423,  429,
+  435,   441,  447,  454,  461,  467,  475,  482,
+  489,   497,  505,  513,  522,  530,  539,  549,
+  559,   569,  579,  590,  602,  614,  626,  640,
+  654,   668,  684,  700,  717,  736,  755,  775,
+  796,   819,  843,  869,  896,  925,  955,  988,
+  1022, 1058, 1098, 1139, 1184, 1232, 1282, 1336,
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const int16_t dc_qlookup_10[QINDEX_RANGE] = {
+  4,     9,    10,    13,    15,    17,    20,    22,
+  25,    28,    31,    34,    37,    40,    43,    47,
+  50,    53,    57,    60,    64,    68,    71,    75,
+  78,    82,    86,    90,    93,    97,   101,   105,
+  109,   113,   116,   120,   124,   128,   132,   136,
+  140,   143,   147,   151,   155,   159,   163,   166,
+  170,   174,   178,   182,   185,   189,   193,   197,
+  200,   204,   208,   212,   215,   219,   223,   226,
+  230,   233,   237,   241,   244,   248,   251,   255,
+  259,   262,   266,   269,   273,   276,   280,   283,
+  287,   290,   293,   297,   300,   304,   307,   310,
+  314,   317,   321,   324,   327,   331,   334,   337,
+  343,   350,   356,   362,   369,   375,   381,   387,
+  394,   400,   406,   412,   418,   424,   430,   436,
+  442,   448,   454,   460,   466,   472,   478,   484,
+  490,   499,   507,   516,   525,   533,   542,   550,
+  559,   567,   576,   584,   592,   601,   609,   617,
+  625,   634,   644,   655,   666,   676,   687,   698,
+  708,   718,   729,   739,   749,   759,   770,   782,
+  795,   807,   819,   831,   844,   856,   868,   880,
+  891,   906,   920,   933,   947,   961,   975,   988,
+  1001,  1015,  1030,  1045,  1061,  1076,  1090,  1105,
+  1120,  1137,  1153,  1170,  1186,  1202,  1218,  1236,
+  1253,  1271,  1288,  1306,  1323,  1342,  1361,  1379,
+  1398,  1416,  1436,  1456,  1476,  1496,  1516,  1537,
+  1559,  1580,  1601,  1624,  1647,  1670,  1692,  1717,
+  1741,  1766,  1791,  1817,  1844,  1871,  1900,  1929,
+  1958,  1990,  2021,  2054,  2088,  2123,  2159,  2197,
+  2236,  2276,  2319,  2363,  2410,  2458,  2508,  2561,
+  2616,  2675,  2737,  2802,  2871,  2944,  3020,  3102,
+  3188,  3280,  3375,  3478,  3586,  3702,  3823,  3953,
+  4089,  4236,  4394,  4559,  4737,  4929,  5130,  5347,
+};
+
+static const int16_t dc_qlookup_12[QINDEX_RANGE] = {
+  4,    12,    18,    25,    33,    41,    50,    60,
+  70,    80,    91,   103,   115,   127,   140,   153,
+  166,   180,   194,   208,   222,   237,   251,   266,
+  281,   296,   312,   327,   343,   358,   374,   390,
+  405,   421,   437,   453,   469,   484,   500,   516,
+  532,   548,   564,   580,   596,   611,   627,   643,
+  659,   674,   690,   706,   721,   737,   752,   768,
+  783,   798,   814,   829,   844,   859,   874,   889,
+  904,   919,   934,   949,   964,   978,   993,  1008,
+  1022,  1037,  1051,  1065,  1080,  1094,  1108,  1122,
+  1136,  1151,  1165,  1179,  1192,  1206,  1220,  1234,
+  1248,  1261,  1275,  1288,  1302,  1315,  1329,  1342,
+  1368,  1393,  1419,  1444,  1469,  1494,  1519,  1544,
+  1569,  1594,  1618,  1643,  1668,  1692,  1717,  1741,
+  1765,  1789,  1814,  1838,  1862,  1885,  1909,  1933,
+  1957,  1992,  2027,  2061,  2096,  2130,  2165,  2199,
+  2233,  2267,  2300,  2334,  2367,  2400,  2434,  2467,
+  2499,  2532,  2575,  2618,  2661,  2704,  2746,  2788,
+  2830,  2872,  2913,  2954,  2995,  3036,  3076,  3127,
+  3177,  3226,  3275,  3324,  3373,  3421,  3469,  3517,
+  3565,  3621,  3677,  3733,  3788,  3843,  3897,  3951,
+  4005,  4058,  4119,  4181,  4241,  4301,  4361,  4420,
+  4479,  4546,  4612,  4677,  4742,  4807,  4871,  4942,
+  5013,  5083,  5153,  5222,  5291,  5367,  5442,  5517,
+  5591,  5665,  5745,  5825,  5905,  5984,  6063,  6149,
+  6234,  6319,  6404,  6495,  6587,  6678,  6769,  6867,
+  6966,  7064,  7163,  7269,  7376,  7483,  7599,  7715,
+  7832,  7958,  8085,  8214,  8352,  8492,  8635,  8788,
+  8945,  9104,  9275,  9450,  9639,  9832, 10031, 10245,
+  10465, 10702, 10946, 11210, 11482, 11776, 12081, 12409,
+  12750, 13118, 13501, 13913, 14343, 14807, 15290, 15812,
+  16356, 16943, 17575, 18237, 18949, 19718, 20521, 21387,
+};
+#endif
+
+static const int16_t ac_qlookup[QINDEX_RANGE] = {
+  4,       8,    9,   10,   11,   12,   13,   14,
+  15,     16,   17,   18,   19,   20,   21,   22,
+  23,     24,   25,   26,   27,   28,   29,   30,
+  31,     32,   33,   34,   35,   36,   37,   38,
+  39,     40,   41,   42,   43,   44,   45,   46,
+  47,     48,   49,   50,   51,   52,   53,   54,
+  55,     56,   57,   58,   59,   60,   61,   62,
+  63,     64,   65,   66,   67,   68,   69,   70,
+  71,     72,   73,   74,   75,   76,   77,   78,
+  79,     80,   81,   82,   83,   84,   85,   86,
+  87,     88,   89,   90,   91,   92,   93,   94,
+  95,     96,   97,   98,   99,  100,  101,  102,
+  104,   106,  108,  110,  112,  114,  116,  118,
+  120,   122,  124,  126,  128,  130,  132,  134,
+  136,   138,  140,  142,  144,  146,  148,  150,
+  152,   155,  158,  161,  164,  167,  170,  173,
+  176,   179,  182,  185,  188,  191,  194,  197,
+  200,   203,  207,  211,  215,  219,  223,  227,
+  231,   235,  239,  243,  247,  251,  255,  260,
+  265,   270,  275,  280,  285,  290,  295,  300,
+  305,   311,  317,  323,  329,  335,  341,  347,
+  353,   359,  366,  373,  380,  387,  394,  401,
+  408,   416,  424,  432,  440,  448,  456,  465,
+  474,   483,  492,  501,  510,  520,  530,  540,
+  550,   560,  571,  582,  593,  604,  615,  627,
+  639,   651,  663,  676,  689,  702,  715,  729,
+  743,   757,  771,  786,  801,  816,  832,  848,
+  864,   881,  898,  915,  933,  951,  969,  988,
+  1007, 1026, 1046, 1066, 1087, 1108, 1129, 1151,
+  1173, 1196, 1219, 1243, 1267, 1292, 1317, 1343,
+  1369, 1396, 1423, 1451, 1479, 1508, 1537, 1567,
+  1597, 1628, 1660, 1692, 1725, 1759, 1793, 1828,
+};
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static const int16_t ac_qlookup_10[QINDEX_RANGE] = {
+  4,     9,    11,    13,    16,    18,    21,    24,
+  27,    30,    33,    37,    40,    44,    48,    51,
+  55,    59,    63,    67,    71,    75,    79,    83,
+  88,    92,    96,   100,   105,   109,   114,   118,
+  122,   127,   131,   136,   140,   145,   149,   154,
+  158,   163,   168,   172,   177,   181,   186,   190,
+  195,   199,   204,   208,   213,   217,   222,   226,
+  231,   235,   240,   244,   249,   253,   258,   262,
+  267,   271,   275,   280,   284,   289,   293,   297,
+  302,   306,   311,   315,   319,   324,   328,   332,
+  337,   341,   345,   349,   354,   358,   362,   367,
+  371,   375,   379,   384,   388,   392,   396,   401,
+  409,   417,   425,   433,   441,   449,   458,   466,
+  474,   482,   490,   498,   506,   514,   523,   531,
+  539,   547,   555,   563,   571,   579,   588,   596,
+  604,   616,   628,   640,   652,   664,   676,   688,
+  700,   713,   725,   737,   749,   761,   773,   785,
+  797,   809,   825,   841,   857,   873,   889,   905,
+  922,   938,   954,   970,   986,  1002,  1018,  1038,
+  1058,  1078,  1098,  1118,  1138,  1158,  1178,  1198,
+  1218,  1242,  1266,  1290,  1314,  1338,  1362,  1386,
+  1411,  1435,  1463,  1491,  1519,  1547,  1575,  1603,
+  1631,  1663,  1695,  1727,  1759,  1791,  1823,  1859,
+  1895,  1931,  1967,  2003,  2039,  2079,  2119,  2159,
+  2199,  2239,  2283,  2327,  2371,  2415,  2459,  2507,
+  2555,  2603,  2651,  2703,  2755,  2807,  2859,  2915,
+  2971,  3027,  3083,  3143,  3203,  3263,  3327,  3391,
+  3455,  3523,  3591,  3659,  3731,  3803,  3876,  3952,
+  4028,  4104,  4184,  4264,  4348,  4432,  4516,  4604,
+  4692,  4784,  4876,  4972,  5068,  5168,  5268,  5372,
+  5476,  5584,  5692,  5804,  5916,  6032,  6148,  6268,
+  6388,  6512,  6640,  6768,  6900,  7036,  7172,  7312,
+};
+
+static const int16_t ac_qlookup_12[QINDEX_RANGE] = {
+  4,    13,    19,    27,    35,    44,    54,    64,
+  75,    87,    99,   112,   126,   139,   154,   168,
+  183,   199,   214,   230,   247,   263,   280,   297,
+  314,   331,   349,   366,   384,   402,   420,   438,
+  456,   475,   493,   511,   530,   548,   567,   586,
+  604,   623,   642,   660,   679,   698,   716,   735,
+  753,   772,   791,   809,   828,   846,   865,   884,
+  902,   920,   939,   957,   976,   994,  1012,  1030,
+  1049,  1067,  1085,  1103,  1121,  1139,  1157,  1175,
+  1193,  1211,  1229,  1246,  1264,  1282,  1299,  1317,
+  1335,  1352,  1370,  1387,  1405,  1422,  1440,  1457,
+  1474,  1491,  1509,  1526,  1543,  1560,  1577,  1595,
+  1627,  1660,  1693,  1725,  1758,  1791,  1824,  1856,
+  1889,  1922,  1954,  1987,  2020,  2052,  2085,  2118,
+  2150,  2183,  2216,  2248,  2281,  2313,  2346,  2378,
+  2411,  2459,  2508,  2556,  2605,  2653,  2701,  2750,
+  2798,  2847,  2895,  2943,  2992,  3040,  3088,  3137,
+  3185,  3234,  3298,  3362,  3426,  3491,  3555,  3619,
+  3684,  3748,  3812,  3876,  3941,  4005,  4069,  4149,
+  4230,  4310,  4390,  4470,  4550,  4631,  4711,  4791,
+  4871,  4967,  5064,  5160,  5256,  5352,  5448,  5544,
+  5641,  5737,  5849,  5961,  6073,  6185,  6297,  6410,
+  6522,  6650,  6778,  6906,  7034,  7162,  7290,  7435,
+  7579,  7723,  7867,  8011,  8155,  8315,  8475,  8635,
+  8795,  8956,  9132,  9308,  9484,  9660,  9836, 10028,
+  10220, 10412, 10604, 10812, 11020, 11228, 11437, 11661,
+  11885, 12109, 12333, 12573, 12813, 13053, 13309, 13565,
+  13821, 14093, 14365, 14637, 14925, 15213, 15502, 15806,
+  16110, 16414, 16734, 17054, 17390, 17726, 18062, 18414,
+  18766, 19134, 19502, 19886, 20270, 20670, 21070, 21486,
+  21902, 22334, 22766, 23214, 23662, 24126, 24590, 25070,
+  25551, 26047, 26559, 27071, 27599, 28143, 28687, 29247,
+};
+#endif
+
+int16_t vp9_dc_quant(int qindex, int delta, vpx_bit_depth_t bit_depth) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_10:
+      return dc_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_12:
+      return dc_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int16_t vp9_ac_quant(int qindex, int delta, vpx_bit_depth_t bit_depth) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  switch (bit_depth) {
+    case VPX_BITS_8:
+      return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_10:
+      return ac_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case VPX_BITS_12:
+      return ac_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
+      return -1;
+  }
+#else
+  (void) bit_depth;
+  return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int vp9_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex) {
+  if (vp9_segfeature_active(seg, segment_id, SEG_LVL_ALT_Q)) {
+    const int data = vp9_get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
+    const int seg_qindex = seg->abs_delta == SEGMENT_ABSDATA ?
+        data : base_qindex + data;
+    return clamp(seg_qindex, 0, MAXQ);
+  } else {
+    return base_qindex;
+  }
+}
+
diff --git a/media/libvpx/vp9/common/vp9_quant_common.h b/media/libvpx/vp9/common/vp9_quant_common.h
new file mode 100644
index 000000000..4bae4a896
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_quant_common.h
@@ -0,0 +1,36 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_QUANT_COMMON_H_
+#define VP9_COMMON_VP9_QUANT_COMMON_H_
+
+#include "vpx/vpx_codec.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MINQ 0
+#define MAXQ 255
+#define QINDEX_RANGE (MAXQ - MINQ + 1)
+#define QINDEX_BITS 8
+
+int16_t vp9_dc_quant(int qindex, int delta, vpx_bit_depth_t bit_depth);
+int16_t vp9_ac_quant(int qindex, int delta, vpx_bit_depth_t bit_depth);
+
+int vp9_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_QUANT_COMMON_H_
diff --git a/media/libvpx/vp9/common/vp9_reconinter.c b/media/libvpx/vp9/common/vp9_reconinter.c
new file mode 100644
index 000000000..11eaf2e2d
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_reconinter.c
@@ -0,0 +1,317 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "./vpx_scale_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_reconintra.h"
+
+void inter_predictor(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            const int subpel_x,
+                            const int subpel_y,
+                            const struct scale_factors *sf,
+                            int w, int h, int ref,
+                            const InterpKernel *kernel,
+                            int xs, int ys) {
+  sf->predict[subpel_x != 0][subpel_y != 0][ref](
+      src, src_stride, dst, dst_stride,
+      kernel[subpel_x], xs, kernel[subpel_y], ys, w, h);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void high_inter_predictor(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 const int subpel_x,
+                                 const int subpel_y,
+                                 const struct scale_factors *sf,
+                                 int w, int h, int ref,
+                                 const InterpKernel *kernel,
+                                 int xs, int ys, int bd) {
+  sf->highbd_predict[subpel_x != 0][subpel_y != 0][ref](
+      src, src_stride, dst, dst_stride,
+      kernel[subpel_x], xs, kernel[subpel_y], ys, w, h, bd);
+}
+
+void vp9_highbd_build_inter_predictor(const uint8_t *src, int src_stride,
+                                      uint8_t *dst, int dst_stride,
+                                      const MV *src_mv,
+                                      const struct scale_factors *sf,
+                                      int w, int h, int ref,
+                                      const InterpKernel *kernel,
+                                      enum mv_precision precision,
+                                      int x, int y, int bd) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  high_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                       sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride,
+                               const MV *src_mv,
+                               const struct scale_factors *sf,
+                               int w, int h, int ref,
+                               const InterpKernel *kernel,
+                               enum mv_precision precision,
+                               int x, int y) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                  sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4);
+}
+
+static INLINE int round_mv_comp_q4(int value) {
+  return (value < 0 ? value - 2 : value + 2) / 4;
+}
+
+static MV mi_mv_pred_q4(const MODE_INFO *mi, int idx) {
+  MV res = { round_mv_comp_q4(mi->bmi[0].as_mv[idx].as_mv.row +
+                              mi->bmi[1].as_mv[idx].as_mv.row +
+                              mi->bmi[2].as_mv[idx].as_mv.row +
+                              mi->bmi[3].as_mv[idx].as_mv.row),
+             round_mv_comp_q4(mi->bmi[0].as_mv[idx].as_mv.col +
+                              mi->bmi[1].as_mv[idx].as_mv.col +
+                              mi->bmi[2].as_mv[idx].as_mv.col +
+                              mi->bmi[3].as_mv[idx].as_mv.col) };
+  return res;
+}
+
+static INLINE int round_mv_comp_q2(int value) {
+  return (value < 0 ? value - 1 : value + 1) / 2;
+}
+
+static MV mi_mv_pred_q2(const MODE_INFO *mi, int idx, int block0, int block1) {
+  MV res = { round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.row +
+                              mi->bmi[block1].as_mv[idx].as_mv.row),
+             round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.col +
+                              mi->bmi[block1].as_mv[idx].as_mv.col) };
+  return res;
+}
+
+// TODO(jkoleszar): yet another mv clamping function :-(
+MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
+                             int bw, int bh, int ss_x, int ss_y) {
+  // If the MV points so far into the UMV border that no visible pixels
+  // are used for reconstruction, the subpel part of the MV can be
+  // discarded and the MV limited to 16 pixels with equivalent results.
+  const int spel_left = (VP9_INTERP_EXTEND + bw) << SUBPEL_BITS;
+  const int spel_right = spel_left - SUBPEL_SHIFTS;
+  const int spel_top = (VP9_INTERP_EXTEND + bh) << SUBPEL_BITS;
+  const int spel_bottom = spel_top - SUBPEL_SHIFTS;
+  MV clamped_mv = {
+    src_mv->row * (1 << (1 - ss_y)),
+    src_mv->col * (1 << (1 - ss_x))
+  };
+  assert(ss_x <= 1);
+  assert(ss_y <= 1);
+
+  clamp_mv(&clamped_mv,
+           xd->mb_to_left_edge * (1 << (1 - ss_x)) - spel_left,
+           xd->mb_to_right_edge * (1 << (1 - ss_x)) + spel_right,
+           xd->mb_to_top_edge * (1 << (1 - ss_y)) - spel_top,
+           xd->mb_to_bottom_edge * (1 << (1 - ss_y)) + spel_bottom);
+
+  return clamped_mv;
+}
+
+MV average_split_mvs(const struct macroblockd_plane *pd,
+                     const MODE_INFO *mi, int ref, int block) {
+  const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
+  MV res = {0, 0};
+  switch (ss_idx) {
+    case 0:
+      res = mi->bmi[block].as_mv[ref].as_mv;
+      break;
+    case 1:
+      res = mi_mv_pred_q2(mi, ref, block, block + 2);
+      break;
+    case 2:
+      res = mi_mv_pred_q2(mi, ref, block, block + 1);
+      break;
+    case 3:
+      res = mi_mv_pred_q4(mi, ref);
+      break;
+    default:
+      assert(ss_idx <= 3 && ss_idx >= 0);
+  }
+  return res;
+}
+
+void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                   int bw, int bh,
+                                   int x, int y, int w, int h,
+                                   int mi_x, int mi_y) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO *mi = xd->mi[0];
+  const int is_compound = has_second_ref(&mi->mbmi);
+  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
+  int ref;
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+    struct buf_2d *const pre_buf = &pd->pre[ref];
+    struct buf_2d *const dst_buf = &pd->dst;
+    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
+               ? average_split_mvs(pd, mi, ref, block)
+               : mi->mbmi.mv[ref].as_mv;
+
+    // TODO(jkoleszar): This clamping is done in the incorrect place for the
+    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+    // MV. Note however that it performs the subsampling aware scaling so
+    // that the result is always q4.
+    // mv_precision precision is MV_PRECISION_Q4.
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+
+    uint8_t *pre;
+    MV32 scaled_mv;
+    int xs, ys, subpel_x, subpel_y;
+    const int is_scaled = vp9_is_scaled(sf);
+
+    if (is_scaled) {
+      pre = pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
+      scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+      xs = sf->x_step_q4;
+      ys = sf->y_step_q4;
+    } else {
+      pre = pre_buf->buf + (y * pre_buf->stride + x);
+      scaled_mv.row = mv_q4.row;
+      scaled_mv.col = mv_q4.col;
+      xs = ys = 16;
+    }
+    subpel_x = scaled_mv.col & SUBPEL_MASK;
+    subpel_y = scaled_mv.row & SUBPEL_MASK;
+    pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride
+           + (scaled_mv.col >> SUBPEL_BITS);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      high_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                           subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys,
+                           xd->bd);
+    } else {
+      inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                      subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+    }
+#else
+    inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                    subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+  }
+}
+
+static void build_inter_predictors_for_planes(MACROBLOCKD *xd, BLOCK_SIZE bsize,
+                                              int mi_row, int mi_col,
+                                              int plane_from, int plane_to) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) {
+      int i = 0, x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+           build_inter_predictors(xd, plane, i++, bw, bh,
+                                  4 * x, 4 * y, 4, 4, mi_x, mi_y);
+    } else {
+      build_inter_predictors(xd, plane, 0, bw, bh,
+                             0, 0, bw, bh, mi_x, mi_y);
+    }
+  }
+}
+
+void vp9_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0, 0);
+}
+
+void vp9_build_inter_predictors_sbp(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize, int plane) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, plane, plane);
+}
+
+void vp9_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 1,
+                                    MAX_MB_PLANE - 1);
+}
+
+void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0,
+                                    MAX_MB_PLANE - 1);
+}
+
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col) {
+  uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+      src->v_buffer};
+  const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+      src->uv_stride};
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &planes[i];
+    setup_pred_plane(&pd->dst, buffers[i], strides[i], mi_row, mi_col, NULL,
+                     pd->subsampling_x, pd->subsampling_y);
+  }
+}
+
+void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col,
+                          const struct scale_factors *sf) {
+  if (src != NULL) {
+    int i;
+    uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+        src->v_buffer};
+    const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+        src->uv_stride};
+    for (i = 0; i < MAX_MB_PLANE; ++i) {
+      struct macroblockd_plane *const pd = &xd->plane[i];
+      setup_pred_plane(&pd->pre[idx], buffers[i], strides[i], mi_row, mi_col,
+                       sf, pd->subsampling_x, pd->subsampling_y);
+    }
+  }
+}
diff --git a/media/libvpx/vp9/common/vp9_reconinter.h b/media/libvpx/vp9/common/vp9_reconinter.h
new file mode 100644
index 000000000..e7057445a
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_reconinter.h
@@ -0,0 +1,114 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_RECONINTER_H_
+#define VP9_COMMON_VP9_RECONINTER_H_
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void inter_predictor(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            const int subpel_x,
+                            const int subpel_y,
+                            const struct scale_factors *sf,
+                            int w, int h, int ref,
+                            const InterpKernel *kernel,
+                            int xs, int ys);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void high_inter_predictor(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 const int subpel_x,
+                                 const int subpel_y,
+                                 const struct scale_factors *sf,
+                                 int w, int h, int ref,
+                                 const InterpKernel *kernel,
+                                 int xs, int ys, int bd);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+MV average_split_mvs(const struct macroblockd_plane *pd, const MODE_INFO *mi,
+                     int ref, int block);
+
+MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
+                             int bw, int bh, int ss_x, int ss_y);
+
+void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                   int bw, int bh,
+                                   int x, int y, int w, int h,
+                                   int mi_x, int mi_y);
+
+void vp9_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictors_sbp(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BLOCK_SIZE bsize, int plane);
+
+void vp9_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize);
+
+void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
+                               uint8_t *dst, int dst_stride,
+                               const MV *mv_q3,
+                               const struct scale_factors *sf,
+                               int w, int h, int do_avg,
+                               const InterpKernel *kernel,
+                               enum mv_precision precision,
+                               int x, int y);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_highbd_build_inter_predictor(const uint8_t *src, int src_stride,
+                                      uint8_t *dst, int dst_stride,
+                                      const MV *mv_q3,
+                                      const struct scale_factors *sf,
+                                      int w, int h, int do_avg,
+                                      const InterpKernel *kernel,
+                                      enum mv_precision precision,
+                                      int x, int y, int bd);
+#endif
+
+static INLINE int scaled_buffer_offset(int x_offset, int y_offset, int stride,
+                                       const struct scale_factors *sf) {
+  const int x = sf ? sf->scale_value_x(x_offset, sf) : x_offset;
+  const int y = sf ? sf->scale_value_y(y_offset, sf) : y_offset;
+  return y * stride + x;
+}
+
+static INLINE void setup_pred_plane(struct buf_2d *dst,
+                                    uint8_t *src, int stride,
+                                    int mi_row, int mi_col,
+                                    const struct scale_factors *scale,
+                                    int subsampling_x, int subsampling_y) {
+  const int x = (MI_SIZE * mi_col) >> subsampling_x;
+  const int y = (MI_SIZE * mi_row) >> subsampling_y;
+  dst->buf = src + scaled_buffer_offset(x, y, stride, scale);
+  dst->stride = stride;
+}
+
+void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col);
+
+void vp9_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *sf);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_RECONINTER_H_
diff --git a/media/libvpx/vp9/common/vp9_reconintra.c b/media/libvpx/vp9/common/vp9_reconintra.c
new file mode 100644
index 000000000..3312f2977
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_reconintra.c
@@ -0,0 +1,1088 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/vpx_once.h"
+
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_onyxc_int.h"
+
+const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES] = {
+  DCT_DCT,    // DC
+  ADST_DCT,   // V
+  DCT_ADST,   // H
+  DCT_DCT,    // D45
+  ADST_ADST,  // D135
+  ADST_DCT,   // D117
+  DCT_ADST,   // D153
+  DCT_ADST,   // D207
+  ADST_DCT,   // D63
+  ADST_ADST,  // TM
+};
+
+enum {
+  NEED_LEFT = 1 << 1,
+  NEED_ABOVE = 1 << 2,
+  NEED_ABOVERIGHT = 1 << 3,
+};
+
+static const uint8_t extend_modes[INTRA_MODES] = {
+  NEED_ABOVE | NEED_LEFT,       // DC
+  NEED_ABOVE,                   // V
+  NEED_LEFT,                    // H
+  NEED_ABOVERIGHT,              // D45
+  NEED_LEFT | NEED_ABOVE,       // D135
+  NEED_LEFT | NEED_ABOVE,       // D117
+  NEED_LEFT | NEED_ABOVE,       // D153
+  NEED_LEFT,                    // D207
+  NEED_ABOVERIGHT,              // D63
+  NEED_LEFT | NEED_ABOVE,       // TM
+};
+
+// This serves as a wrapper function, so that all the prediction functions
+// can be unified and accessed as a pointer array. Note that the boundary
+// above and left are not necessarily used all the time.
+#define intra_pred_sized(type, size) \
+  void vp9_##type##_predictor_##size##x##size##_c(uint8_t *dst, \
+                                                  ptrdiff_t stride, \
+                                                  const uint8_t *above, \
+                                                  const uint8_t *left) { \
+    type##_predictor(dst, stride, size, above, left); \
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+#define intra_pred_highbd_sized(type, size) \
+  void vp9_highbd_##type##_predictor_##size##x##size##_c( \
+      uint16_t *dst, ptrdiff_t stride, const uint16_t *above, \
+      const uint16_t *left, int bd) { \
+    highbd_##type##_predictor(dst, stride, size, above, left, bd); \
+  }
+
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+
+#define intra_pred_no_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+
+#else
+
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+
+#define intra_pred_no_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#define DST(x, y) dst[(x) + (y) * stride]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG2(a, b) (((a) + (b) + 1) >> 1)
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE void highbd_d207_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) above;
+  (void) bd;
+
+  // First column.
+  for (r = 0; r < bs - 1; ++r) {
+    dst[r * stride] = AVG2(left[r], left[r + 1]);
+  }
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // Second column.
+  for (r = 0; r < bs - 2; ++r) {
+    dst[r * stride] = AVG3(left[r], left[r + 1], left[r + 2]);
+  }
+  dst[(bs - 2) * stride] = AVG3(left[bs - 2], left[bs - 1], left[bs - 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // Rest of last row.
+  for (c = 0; c < bs - 2; ++c)
+    dst[(bs - 1) * stride + c] = left[bs - 1];
+
+  for (r = bs - 2; r >= 0; --r) {
+    for (c = 0; c < bs - 2; ++c)
+      dst[r * stride + c] = dst[(r + 1) * stride + c - 2];
+  }
+}
+
+static INLINE void highbd_d63_predictor(uint16_t *dst, ptrdiff_t stride,
+                                        int bs, const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  int r, c;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r & 1 ? AVG3(above[(r >> 1) + c], above[(r >> 1) + c + 1],
+                            above[(r >> 1) + c + 2])
+                     : AVG2(above[(r >> 1) + c], above[(r >> 1) + c + 1]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d45_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  int r, c;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r + c + 2 < bs * 2 ? AVG3(above[r + c], above[r + c + 1],
+                                         above[r + c + 2])
+                                  : above[bs * 2 - 1];
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d117_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+
+  // first row
+  for (c = 0; c < bs; c++)
+    dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d135_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; ++r)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+
+  dst += stride;
+  for (r = 1; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d153_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void) bd;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++)
+    dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++)
+      dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_v_predictor(uint16_t *dst, ptrdiff_t stride,
+                                      int bs, const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void) left;
+  (void) bd;
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs * sizeof(uint16_t));
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_h_predictor(uint16_t *dst, ptrdiff_t stride,
+                                      int bs, const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void) above;
+  (void) bd;
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_tm_predictor(uint16_t *dst, ptrdiff_t stride,
+                                       int bs, const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int r, c;
+  int ytop_left = above[-1];
+  (void) bd;
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel_highbd(left[r] + above[c] - ytop_left, bd);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_128_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int r;
+  (void) above;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, 128 << (bd - 8), bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_left_predictor(uint16_t *dst, ptrdiff_t stride,
+                                            int bs, const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void) above;
+  (void) bd;
+
+  for (i = 0; i < bs; i++)
+    sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_top_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void) left;
+  (void) bd;
+
+  for (i = 0; i < bs; i++)
+    sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride,
+                                       int bs, const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+  (void) bd;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    vpx_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+void vp9_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  (void)above;
+  DST(0, 0) =             AVG2(I, J);
+  DST(2, 0) = DST(0, 1) = AVG2(J, K);
+  DST(2, 1) = DST(0, 2) = AVG2(K, L);
+  DST(1, 0) =             AVG3(I, J, K);
+  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
+  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
+  DST(3, 2) = DST(2, 2) =
+      DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
+}
+
+static INLINE void d207_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void) above;
+  // first column
+  for (r = 0; r < bs - 1; ++r)
+    dst[r * stride] = AVG2(left[r], left[r + 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // second column
+  for (r = 0; r < bs - 2; ++r)
+    dst[r * stride] = AVG3(left[r], left[r + 1], left[r + 2]);
+  dst[(bs - 2) * stride] = AVG3(left[bs - 2], left[bs - 1], left[bs - 1]);
+  dst[(bs - 1) * stride] = left[bs - 1];
+  dst++;
+
+  // rest of last row
+  for (c = 0; c < bs - 2; ++c)
+    dst[(bs - 1) * stride + c] = left[bs - 1];
+
+  for (r = bs - 2; r >= 0; --r)
+    for (c = 0; c < bs - 2; ++c)
+      dst[r * stride + c] = dst[(r + 1) * stride + c - 2];
+}
+intra_pred_no_4x4(d207)
+
+void vp9_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                             const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  (void)left;
+  DST(0, 0) =             AVG2(A, B);
+  DST(1, 0) = DST(0, 2) = AVG2(B, C);
+  DST(2, 0) = DST(1, 2) = AVG2(C, D);
+  DST(3, 0) = DST(2, 2) = AVG2(D, E);
+              DST(3, 2) = AVG2(E, F);  // differs from vp8
+
+  DST(0, 1) =             AVG3(A, B, C);
+  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
+  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
+  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
+              DST(3, 3) = AVG3(E, F, G);  // differs from vp8
+}
+
+static INLINE void d63_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  int size;
+  (void)left;
+  for (c = 0; c < bs; ++c) {
+    dst[c] = AVG2(above[c], above[c + 1]);
+    dst[stride + c] = AVG3(above[c], above[c + 1], above[c + 2]);
+  }
+  for (r = 2, size = bs - 2; r < bs; r += 2, --size) {
+    memcpy(dst + (r + 0) * stride, dst + (r >> 1), size);
+    memset(dst + (r + 0) * stride + size, above[bs - 1], bs - size);
+    memcpy(dst + (r + 1) * stride, dst + stride + (r >> 1), size);
+    memset(dst + (r + 1) * stride + size, above[bs - 1], bs - size);
+  }
+}
+intra_pred_no_4x4(d63)
+
+void vp9_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                             const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  const int H = above[7];
+  (void)stride;
+  (void)left;
+  DST(0, 0)                                     = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
+  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
+  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
+              DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G);
+                          DST(3, 2) = DST(2, 3) = AVG3(F, G, H);
+                                      DST(3, 3) = H;  // differs from vp8
+}
+
+static INLINE void d45_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8_t above_right = above[bs - 1];
+  int x, size;
+  uint8_t avg[31];  // TODO(jzern): this could be block size specific
+  (void)left;
+
+  for (x = 0; x < bs - 1; ++x) {
+    avg[x] = AVG3(above[x], above[x + 1], above[x + 2]);
+  }
+  for (x = 0, size = bs - 1; x < bs; ++x, --size) {
+    memcpy(dst, avg + x, size);
+    memset(dst + size, above_right, x + 1);
+    dst += stride;
+  }
+}
+intra_pred_no_4x4(d45)
+
+void vp9_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  DST(0, 0) = DST(1, 2) = AVG2(X, A);
+  DST(1, 0) = DST(2, 2) = AVG2(A, B);
+  DST(2, 0) = DST(3, 2) = AVG2(B, C);
+  DST(3, 0)             = AVG2(C, D);
+
+  DST(0, 3) =             AVG3(K, J, I);
+  DST(0, 2) =             AVG3(J, I, X);
+  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
+  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
+  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
+  DST(3, 1) =             AVG3(B, C, D);
+}
+
+static INLINE void d117_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+
+  // first row
+  for (c = 0; c < bs; c++)
+    dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+intra_pred_no_4x4(d117)
+
+void vp9_d135_predictor_4x4(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)stride;
+  DST(0, 3)                                     = AVG3(J, K, L);
+  DST(1, 3) = DST(0, 2)                         = AVG3(I, J, K);
+  DST(2, 3) = DST(1, 2) = DST(0, 1)             = AVG3(X, I, J);
+  DST(3, 3) = DST(2, 2) = DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+              DST(3, 2) = DST(2, 1) = DST(1, 0) = AVG3(B, A, X);
+                          DST(3, 1) = DST(2, 0) = AVG3(C, B, A);
+                                      DST(3, 0) = AVG3(D, C, B);
+}
+
+static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++)
+    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; ++r)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+
+  dst += stride;
+  for (r = 1; r < bs; ++r) {
+    for (c = 1; c < bs; c++)
+      dst[c] = dst[-stride + c - 1];
+    dst += stride;
+  }
+}
+intra_pred_no_4x4(d135)
+
+void vp9_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+
+  DST(0, 0) = DST(2, 1) = AVG2(I, X);
+  DST(0, 1) = DST(2, 2) = AVG2(J, I);
+  DST(0, 2) = DST(2, 3) = AVG2(K, J);
+  DST(0, 3)             = AVG2(L, K);
+
+  DST(3, 0)             = AVG3(A, B, C);
+  DST(2, 0)             = AVG3(X, A, B);
+  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
+  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
+  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
+  DST(1, 3)             = AVG3(L, K, J);
+}
+
+static INLINE void d153_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++)
+    dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++)
+      dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+intra_pred_no_4x4(d153)
+
+static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(v)
+
+static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) above;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, left[r], bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(h)
+
+static INLINE void tm_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  int ytop_left = above[-1];
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel(left[r] + above[c] - ytop_left);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(tm)
+
+static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void) above;
+  (void) left;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, 128, bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(dc_128)
+
+static INLINE void dc_left_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void) above;
+
+  for (i = 0; i < bs; i++)
+    sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(dc_left)
+
+static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void) left;
+
+  for (i = 0; i < bs; i++)
+    sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(dc_top)
+
+static INLINE void dc_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+intra_pred_allsizes(dc)
+#undef intra_pred_allsizes
+
+typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left);
+
+static intra_pred_fn pred[INTRA_MODES][TX_SIZES];
+static intra_pred_fn dc_pred[2][2][TX_SIZES];
+
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*intra_high_pred_fn)(uint16_t *dst, ptrdiff_t stride,
+                                   const uint16_t *above, const uint16_t *left,
+                                   int bd);
+static intra_high_pred_fn pred_high[INTRA_MODES][4];
+static intra_high_pred_fn dc_pred_high[2][2][4];
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void vp9_init_intra_predictors_internal(void) {
+#define INIT_ALL_SIZES(p, type) \
+  p[TX_4X4] = vp9_##type##_predictor_4x4; \
+  p[TX_8X8] = vp9_##type##_predictor_8x8; \
+  p[TX_16X16] = vp9_##type##_predictor_16x16; \
+  p[TX_32X32] = vp9_##type##_predictor_32x32
+
+  INIT_ALL_SIZES(pred[V_PRED], v);
+  INIT_ALL_SIZES(pred[H_PRED], h);
+  INIT_ALL_SIZES(pred[D207_PRED], d207);
+  INIT_ALL_SIZES(pred[D45_PRED], d45);
+  INIT_ALL_SIZES(pred[D63_PRED], d63);
+  INIT_ALL_SIZES(pred[D117_PRED], d117);
+  INIT_ALL_SIZES(pred[D135_PRED], d135);
+  INIT_ALL_SIZES(pred[D153_PRED], d153);
+  INIT_ALL_SIZES(pred[TM_PRED], tm);
+
+  INIT_ALL_SIZES(dc_pred[0][0], dc_128);
+  INIT_ALL_SIZES(dc_pred[0][1], dc_top);
+  INIT_ALL_SIZES(dc_pred[1][0], dc_left);
+  INIT_ALL_SIZES(dc_pred[1][1], dc);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  INIT_ALL_SIZES(pred_high[V_PRED], highbd_v);
+  INIT_ALL_SIZES(pred_high[H_PRED], highbd_h);
+  INIT_ALL_SIZES(pred_high[D207_PRED], highbd_d207);
+  INIT_ALL_SIZES(pred_high[D45_PRED], highbd_d45);
+  INIT_ALL_SIZES(pred_high[D63_PRED], highbd_d63);
+  INIT_ALL_SIZES(pred_high[D117_PRED], highbd_d117);
+  INIT_ALL_SIZES(pred_high[D135_PRED], highbd_d135);
+  INIT_ALL_SIZES(pred_high[D153_PRED], highbd_d153);
+  INIT_ALL_SIZES(pred_high[TM_PRED], highbd_tm);
+
+  INIT_ALL_SIZES(dc_pred_high[0][0], highbd_dc_128);
+  INIT_ALL_SIZES(dc_pred_high[0][1], highbd_dc_top);
+  INIT_ALL_SIZES(dc_pred_high[1][0], highbd_dc_left);
+  INIT_ALL_SIZES(dc_pred_high[1][1], highbd_dc);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#undef intra_pred_allsizes
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void build_intra_predictors_high(const MACROBLOCKD *xd,
+                                        const uint8_t *ref8,
+                                        int ref_stride,
+                                        uint8_t *dst8,
+                                        int dst_stride,
+                                        PREDICTION_MODE mode,
+                                        TX_SIZE tx_size,
+                                        int up_available,
+                                        int left_available,
+                                        int right_available,
+                                        int x, int y,
+                                        int plane, int bd) {
+  int i;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  DECLARE_ALIGNED(16, uint16_t, left_col[32]);
+  DECLARE_ALIGNED(16, uint16_t, above_data[64 + 16]);
+  uint16_t *above_row = above_data + 16;
+  const uint16_t *const_above_row = above_row;
+  const int bs = 4 << tx_size;
+  int frame_width, frame_height;
+  int x0, y0;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  //  int base=128;
+  int base = 128 << (bd - 8);
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+
+  // Get current frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = xd->cur_buf->y_width;
+    frame_height = xd->cur_buf->y_height;
+  } else {
+    frame_width = xd->cur_buf->uv_width;
+    frame_height = xd->cur_buf->uv_height;
+  }
+
+  // Get block position in current frame.
+  x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+  y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+  // left
+  if (left_available) {
+    if (xd->mb_to_bottom_edge < 0) {
+      /* slower path if the block needs border extension */
+      if (y0 + bs <= frame_height) {
+        for (i = 0; i < bs; ++i)
+          left_col[i] = ref[i * ref_stride - 1];
+      } else {
+        const int extend_bottom = frame_height - y0;
+        for (i = 0; i < extend_bottom; ++i)
+          left_col[i] = ref[i * ref_stride - 1];
+        for (; i < bs; ++i)
+          left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1];
+      }
+    } else {
+      /* faster path if the block does not need extension */
+      for (i = 0; i < bs; ++i)
+        left_col[i] = ref[i * ref_stride - 1];
+    }
+  } else {
+    // TODO(Peter): this value should probably change for high bitdepth
+    vpx_memset16(left_col, base + 1, bs);
+  }
+
+  // TODO(hkuang) do not extend 2*bs pixels for all modes.
+  // above
+  if (up_available) {
+    const uint16_t *above_ref = ref - ref_stride;
+    if (xd->mb_to_right_edge < 0) {
+      /* slower path if the block needs border extension */
+      if (x0 + 2 * bs <= frame_width) {
+        if (right_available && bs == 4) {
+          memcpy(above_row, above_ref, 2 * bs * sizeof(uint16_t));
+        } else {
+          memcpy(above_row, above_ref, bs * sizeof(uint16_t));
+          vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+        }
+      } else if (x0 + bs <= frame_width) {
+        const int r = frame_width - x0;
+        if (right_available && bs == 4) {
+          memcpy(above_row, above_ref, r * sizeof(uint16_t));
+          vpx_memset16(above_row + r, above_row[r - 1],
+                       x0 + 2 * bs - frame_width);
+        } else {
+          memcpy(above_row, above_ref, bs * sizeof(uint16_t));
+          vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+        }
+      } else if (x0 <= frame_width) {
+        const int r = frame_width - x0;
+        memcpy(above_row, above_ref, r * sizeof(uint16_t));
+        vpx_memset16(above_row + r, above_row[r - 1],
+                       x0 + 2 * bs - frame_width);
+      }
+      // TODO(Peter) this value should probably change for high bitdepth
+      above_row[-1] = left_available ? above_ref[-1] : (base+1);
+    } else {
+      /* faster path if the block does not need extension */
+      if (bs == 4 && right_available && left_available) {
+        const_above_row = above_ref;
+      } else {
+        memcpy(above_row, above_ref, bs * sizeof(uint16_t));
+        if (bs == 4 && right_available)
+          memcpy(above_row + bs, above_ref + bs, bs * sizeof(uint16_t));
+        else
+          vpx_memset16(above_row + bs, above_row[bs - 1], bs);
+        // TODO(Peter): this value should probably change for high bitdepth
+        above_row[-1] = left_available ? above_ref[-1] : (base+1);
+      }
+    }
+  } else {
+    vpx_memset16(above_row, base - 1, bs * 2);
+    // TODO(Peter): this value should probably change for high bitdepth
+    above_row[-1] = base - 1;
+  }
+
+  // predict
+  if (mode == DC_PRED) {
+    dc_pred_high[left_available][up_available][tx_size](dst, dst_stride,
+                                                        const_above_row,
+                                                        left_col, xd->bd);
+  } else {
+    pred_high[mode][tx_size](dst, dst_stride, const_above_row, left_col,
+                             xd->bd);
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref,
+                                   int ref_stride, uint8_t *dst, int dst_stride,
+                                   PREDICTION_MODE mode, TX_SIZE tx_size,
+                                   int up_available, int left_available,
+                                   int right_available, int x, int y,
+                                   int plane) {
+  int i;
+  DECLARE_ALIGNED(16, uint8_t, left_col[32]);
+  DECLARE_ALIGNED(16, uint8_t, above_data[64 + 16]);
+  uint8_t *above_row = above_data + 16;
+  const uint8_t *const_above_row = above_row;
+  const int bs = 4 << tx_size;
+  int frame_width, frame_height;
+  int x0, y0;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+  // ..
+
+  // Get current frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = xd->cur_buf->y_width;
+    frame_height = xd->cur_buf->y_height;
+  } else {
+    frame_width = xd->cur_buf->uv_width;
+    frame_height = xd->cur_buf->uv_height;
+  }
+
+  // Get block position in current frame.
+  x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+  y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+  // NEED_LEFT
+  if (extend_modes[mode] & NEED_LEFT) {
+    if (left_available) {
+      if (xd->mb_to_bottom_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (y0 + bs <= frame_height) {
+          for (i = 0; i < bs; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+        } else {
+          const int extend_bottom = frame_height - y0;
+          for (i = 0; i < extend_bottom; ++i)
+            left_col[i] = ref[i * ref_stride - 1];
+          for (; i < bs; ++i)
+            left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1];
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        for (i = 0; i < bs; ++i)
+          left_col[i] = ref[i * ref_stride - 1];
+      }
+    } else {
+      memset(left_col, 129, bs);
+    }
+  }
+
+  // NEED_ABOVE
+  if (extend_modes[mode] & NEED_ABOVE) {
+    if (up_available) {
+      const uint8_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + bs <= frame_width) {
+          memcpy(above_row, above_ref, bs);
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r);
+          memset(above_row + r, above_row[r - 1], x0 + bs - frame_width);
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs);
+        }
+      }
+      above_row[-1] = left_available ? above_ref[-1] : 129;
+    } else {
+      memset(above_row, 127, bs);
+      above_row[-1] = 127;
+    }
+  }
+
+  // NEED_ABOVERIGHT
+  if (extend_modes[mode] & NEED_ABOVERIGHT) {
+    if (up_available) {
+      const uint8_t *above_ref = ref - ref_stride;
+      if (xd->mb_to_right_edge < 0) {
+        /* slower path if the block needs border extension */
+        if (x0 + 2 * bs <= frame_width) {
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, 2 * bs);
+          } else {
+            memcpy(above_row, above_ref, bs);
+            memset(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 + bs <= frame_width) {
+          const int r = frame_width - x0;
+          if (right_available && bs == 4) {
+            memcpy(above_row, above_ref, r);
+            memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width);
+          } else {
+            memcpy(above_row, above_ref, bs);
+            memset(above_row + bs, above_row[bs - 1], bs);
+          }
+        } else if (x0 <= frame_width) {
+          const int r = frame_width - x0;
+          memcpy(above_row, above_ref, r);
+          memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width);
+        }
+      } else {
+        /* faster path if the block does not need extension */
+        if (bs == 4 && right_available && left_available) {
+          const_above_row = above_ref;
+        } else {
+          memcpy(above_row, above_ref, bs);
+          if (bs == 4 && right_available)
+            memcpy(above_row + bs, above_ref + bs, bs);
+          else
+            memset(above_row + bs, above_row[bs - 1], bs);
+        }
+      }
+      above_row[-1] = left_available ? above_ref[-1] : 129;
+    } else {
+      memset(above_row, 127, bs * 2);
+      above_row[-1] = 127;
+    }
+  }
+
+  // predict
+  if (mode == DC_PRED) {
+    dc_pred[left_available][up_available][tx_size](dst, dst_stride,
+                                                   const_above_row, left_col);
+  } else {
+    pred[mode][tx_size](dst, dst_stride, const_above_row, left_col);
+  }
+}
+
+void vp9_predict_intra_block(const MACROBLOCKD *xd, int block_idx, int bwl_in,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride,
+                             uint8_t *dst, int dst_stride,
+                             int aoff, int loff, int plane) {
+  const int bwl = bwl_in - tx_size;
+  const int wmask = (1 << bwl) - 1;
+  const int have_top = (block_idx >> bwl) || xd->up_available;
+  const int have_left = (block_idx & wmask) || xd->left_available;
+  const int have_right = ((block_idx & wmask) != wmask);
+  const int x = aoff * 4;
+  const int y = loff * 4;
+
+  assert(bwl >= 0);
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    build_intra_predictors_high(xd, ref, ref_stride, dst, dst_stride, mode,
+                                tx_size, have_top, have_left, have_right,
+                                x, y, plane, xd->bd);
+    return;
+  }
+#endif
+  build_intra_predictors(xd, ref, ref_stride, dst, dst_stride, mode, tx_size,
+                         have_top, have_left, have_right, x, y, plane);
+}
+
+void vp9_init_intra_predictors(void) {
+  once(vp9_init_intra_predictors_internal);
+}
diff --git a/media/libvpx/vp9/common/vp9_reconintra.h b/media/libvpx/vp9/common/vp9_reconintra.h
new file mode 100644
index 000000000..da5e435b1
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_reconintra.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_RECONINTRA_H_
+#define VP9_COMMON_VP9_RECONINTRA_H_
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_init_intra_predictors(void);
+
+void vp9_predict_intra_block(const MACROBLOCKD *xd, int block_idx, int bwl_in,
+                             TX_SIZE tx_size, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride,
+                             uint8_t *dst, int dst_stride,
+                             int aoff, int loff, int plane);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_RECONINTRA_H_
diff --git a/media/libvpx/vp9/common/vp9_rtcd.c b/media/libvpx/vp9/common/vp9_rtcd.c
new file mode 100644
index 000000000..2dfa09f50
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_rtcd.c
@@ -0,0 +1,19 @@
+/*
+ *  Copyright (c) 2011 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#define RTCD_C
+#include "./vp9_rtcd.h"
+#include "vpx_ports/vpx_once.h"
+
+void vp9_rtcd() {
+    // TODO(JBB): Remove this once, by insuring that both the encoder and
+    // decoder setup functions are protected by once();
+    once(setup_rtcd_internal);
+}
diff --git a/media/libvpx/vp9/common/vp9_scale.c b/media/libvpx/vp9/common/vp9_scale.c
new file mode 100644
index 000000000..6db8f9caa
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_scale.c
@@ -0,0 +1,165 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_filter.h"
+#include "vp9/common/vp9_scale.h"
+
+static INLINE int scaled_x(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->x_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static INLINE int scaled_y(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->y_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static int unscaled_value(int val, const struct scale_factors *sf) {
+  (void) sf;
+  return val;
+}
+
+static int get_fixed_point_scale_factor(int other_size, int this_size) {
+  // Calculate scaling factor once for each reference frame
+  // and use fixed point scaling factors in decoding and encoding routines.
+  // Hardware implementations can calculate scale factor in device driver
+  // and use multiplication and shifting on hardware instead of division.
+  return (other_size << REF_SCALE_SHIFT) / this_size;
+}
+
+MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
+  const int x_off_q4 = scaled_x(x << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const int y_off_q4 = scaled_y(y << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const MV32 res = {
+    scaled_y(mv->row, sf) + y_off_q4,
+    scaled_x(mv->col, sf) + x_off_q4
+  };
+  return res;
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h,
+                                       int use_highbd) {
+#else
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h) {
+#endif
+  if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
+    sf->x_scale_fp = REF_INVALID_SCALE;
+    sf->y_scale_fp = REF_INVALID_SCALE;
+    return;
+  }
+
+  sf->x_scale_fp = get_fixed_point_scale_factor(other_w, this_w);
+  sf->y_scale_fp = get_fixed_point_scale_factor(other_h, this_h);
+  sf->x_step_q4 = scaled_x(16, sf);
+  sf->y_step_q4 = scaled_y(16, sf);
+
+  if (vp9_is_scaled(sf)) {
+    sf->scale_value_x = scaled_x;
+    sf->scale_value_y = scaled_y;
+  } else {
+    sf->scale_value_x = unscaled_value;
+    sf->scale_value_y = unscaled_value;
+  }
+
+  // TODO(agrange): Investigate the best choice of functions to use here
+  // for EIGHTTAP_SMOOTH. Since it is not interpolating, need to choose what
+  // to do at full-pel offsets. The current selection, where the filter is
+  // applied in one direction only, and not at all for 0,0, seems to give the
+  // best quality, but it may be worth trying an additional mode that does
+  // do the filtering on full-pel.
+  if (sf->x_step_q4 == 16) {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in either direction.
+      sf->predict[0][0][0] = vp9_convolve_copy;
+      sf->predict[0][0][1] = vp9_convolve_avg;
+      sf->predict[0][1][0] = vp9_convolve8_vert;
+      sf->predict[0][1][1] = vp9_convolve8_avg_vert;
+      sf->predict[1][0][0] = vp9_convolve8_horiz;
+      sf->predict[1][0][1] = vp9_convolve8_avg_horiz;
+    } else {
+      // No scaling in x direction. Must always scale in the y direction.
+      sf->predict[0][0][0] = vp9_convolve8_vert;
+      sf->predict[0][0][1] = vp9_convolve8_avg_vert;
+      sf->predict[0][1][0] = vp9_convolve8_vert;
+      sf->predict[0][1][1] = vp9_convolve8_avg_vert;
+      sf->predict[1][0][0] = vp9_convolve8;
+      sf->predict[1][0][1] = vp9_convolve8_avg;
+    }
+  } else {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in the y direction. Must always scale in the x direction.
+      sf->predict[0][0][0] = vp9_convolve8_horiz;
+      sf->predict[0][0][1] = vp9_convolve8_avg_horiz;
+      sf->predict[0][1][0] = vp9_convolve8;
+      sf->predict[0][1][1] = vp9_convolve8_avg;
+      sf->predict[1][0][0] = vp9_convolve8_horiz;
+      sf->predict[1][0][1] = vp9_convolve8_avg_horiz;
+    } else {
+      // Must always scale in both directions.
+      sf->predict[0][0][0] = vp9_convolve8;
+      sf->predict[0][0][1] = vp9_convolve8_avg;
+      sf->predict[0][1][0] = vp9_convolve8;
+      sf->predict[0][1][1] = vp9_convolve8_avg;
+      sf->predict[1][0][0] = vp9_convolve8;
+      sf->predict[1][0][1] = vp9_convolve8_avg;
+    }
+  }
+  // 2D subpel motion always gets filtered in both directions
+  sf->predict[1][1][0] = vp9_convolve8;
+  sf->predict[1][1][1] = vp9_convolve8_avg;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (use_highbd) {
+    if (sf->x_step_q4 == 16) {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in either direction.
+        sf->highbd_predict[0][0][0] = vp9_highbd_convolve_copy;
+        sf->highbd_predict[0][0][1] = vp9_highbd_convolve_avg;
+        sf->highbd_predict[0][1][0] = vp9_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = vp9_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = vp9_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = vp9_highbd_convolve8_avg_horiz;
+      } else {
+        // No scaling in x direction. Must always scale in the y direction.
+        sf->highbd_predict[0][0][0] = vp9_highbd_convolve8_vert;
+        sf->highbd_predict[0][0][1] = vp9_highbd_convolve8_avg_vert;
+        sf->highbd_predict[0][1][0] = vp9_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = vp9_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = vp9_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = vp9_highbd_convolve8_avg;
+      }
+    } else {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in the y direction. Must always scale in the x direction.
+        sf->highbd_predict[0][0][0] = vp9_highbd_convolve8_horiz;
+        sf->highbd_predict[0][0][1] = vp9_highbd_convolve8_avg_horiz;
+        sf->highbd_predict[0][1][0] = vp9_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = vp9_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = vp9_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = vp9_highbd_convolve8_avg_horiz;
+      } else {
+        // Must always scale in both directions.
+        sf->highbd_predict[0][0][0] = vp9_highbd_convolve8;
+        sf->highbd_predict[0][0][1] = vp9_highbd_convolve8_avg;
+        sf->highbd_predict[0][1][0] = vp9_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = vp9_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = vp9_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = vp9_highbd_convolve8_avg;
+      }
+    }
+    // 2D subpel motion always gets filtered in both directions.
+    sf->highbd_predict[1][1][0] = vp9_highbd_convolve8;
+    sf->highbd_predict[1][1][1] = vp9_highbd_convolve8_avg;
+  }
+#endif
+}
diff --git a/media/libvpx/vp9/common/vp9_scale.h b/media/libvpx/vp9/common/vp9_scale.h
new file mode 100644
index 000000000..a1601a72f
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_scale.h
@@ -0,0 +1,75 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SCALE_H_
+#define VP9_COMMON_VP9_SCALE_H_
+
+#include "vp9/common/vp9_mv.h"
+#include "vp9/common/vp9_convolve.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define REF_SCALE_SHIFT 14
+#define REF_NO_SCALE (1 << REF_SCALE_SHIFT)
+#define REF_INVALID_SCALE -1
+
+struct scale_factors {
+  int x_scale_fp;   // horizontal fixed point scale factor
+  int y_scale_fp;   // vertical fixed point scale factor
+  int x_step_q4;
+  int y_step_q4;
+
+  int (*scale_value_x)(int val, const struct scale_factors *sf);
+  int (*scale_value_y)(int val, const struct scale_factors *sf);
+
+  convolve_fn_t predict[2][2][2];  // horiz, vert, avg
+#if CONFIG_VP9_HIGHBITDEPTH
+  highbd_convolve_fn_t highbd_predict[2][2][2];  // horiz, vert, avg
+#endif
+};
+
+MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf);
+
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h,
+                                       int use_high);
+#else
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h);
+#endif
+
+static INLINE int vp9_is_valid_scale(const struct scale_factors *sf) {
+  return sf->x_scale_fp != REF_INVALID_SCALE &&
+         sf->y_scale_fp != REF_INVALID_SCALE;
+}
+
+static INLINE int vp9_is_scaled(const struct scale_factors *sf) {
+  return vp9_is_valid_scale(sf) &&
+         (sf->x_scale_fp != REF_NO_SCALE || sf->y_scale_fp != REF_NO_SCALE);
+}
+
+static INLINE int valid_ref_frame_size(int ref_width, int ref_height,
+                                      int this_width, int this_height) {
+  return 2 * this_width >= ref_width &&
+         2 * this_height >= ref_height &&
+         this_width <= 16 * ref_width &&
+         this_height <= 16 * ref_height;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SCALE_H_
diff --git a/media/libvpx/vp9/common/vp9_scan.c b/media/libvpx/vp9/common/vp9_scan.c
new file mode 100644
index 000000000..d6fb8b2d7
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_scan.c
@@ -0,0 +1,727 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_scan.h"
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_4x4[16]) = {
+  0,  4,  1,  5,
+  8,  2, 12,  9,
+  3,  6, 13, 10,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_4x4[16]) = {
+  0,  4,  8,  1,
+  12,  5,  9,  2,
+  13,  6, 10,  3,
+  7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_4x4[16]) = {
+  0,  1,  4,  2,
+  5,  3,  6,  8,
+  9,  7, 12, 10,
+  13, 11, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x8[64]) = {
+  0,  8,  1, 16,  9,  2, 17, 24,
+  10,  3, 18, 25, 32, 11,  4, 26,
+  33, 19, 40, 12, 34, 27,  5, 41,
+  20, 48, 13, 35, 42, 28, 21,  6,
+  49, 56, 36, 43, 29,  7, 14, 50,
+  57, 44, 22, 37, 15, 51, 58, 30,
+  45, 23, 52, 59, 38, 31, 60, 53,
+  46, 39, 61, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_8x8[64]) = {
+  0,  8, 16,  1, 24,  9, 32, 17,
+  2, 40, 25, 10, 33, 18, 48,  3,
+  26, 41, 11, 56, 19, 34,  4, 49,
+  27, 42, 12, 35, 20, 57, 50, 28,
+  5, 43, 13, 36, 58, 51, 21, 44,
+  6, 29, 59, 37, 14, 52, 22,  7,
+  45, 60, 30, 15, 38, 53, 23, 46,
+  31, 61, 39, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_8x8[64]) = {
+  0,  1,  2,  8,  9,  3, 16, 10,
+  4, 17, 11, 24,  5, 18, 25, 12,
+  19, 26, 32,  6, 13, 20, 33, 27,
+  7, 34, 40, 21, 28, 41, 14, 35,
+  48, 42, 29, 36, 49, 22, 43, 15,
+  56, 37, 50, 44, 30, 57, 23, 51,
+  58, 45, 38, 52, 31, 59, 53, 46,
+  60, 39, 61, 47, 54, 55, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x16[256]) = {
+  0, 16, 1, 32, 17, 2, 48, 33, 18, 3, 64, 34, 49, 19, 65, 80,
+  50, 4, 35, 66, 20, 81, 96, 51, 5, 36, 82, 97, 67, 112, 21, 52,
+  98, 37, 83, 113, 6, 68, 128, 53, 22, 99, 114, 84, 7, 129, 38, 69,
+  100, 115, 144, 130, 85, 54, 23, 8, 145, 39, 70, 116, 101, 131, 160, 146,
+  55, 86, 24, 71, 132, 117, 161, 40, 9, 102, 147, 176, 162, 87, 56, 25,
+  133, 118, 177, 148, 72, 103, 41, 163, 10, 192, 178, 88, 57, 134, 149, 119,
+  26, 164, 73, 104, 193, 42, 179, 208, 11, 135, 89, 165, 120, 150, 58, 194,
+  180, 27, 74, 209, 105, 151, 136, 43, 90, 224, 166, 195, 181, 121, 210, 59,
+  12, 152, 106, 167, 196, 75, 137, 225, 211, 240, 182, 122, 91, 28, 197, 13,
+  226, 168, 183, 153, 44, 212, 138, 107, 241, 60, 29, 123, 198, 184, 227, 169,
+  242, 76, 213, 154, 45, 92, 14, 199, 139, 61, 228, 214, 170, 185, 243, 108,
+  77, 155, 30, 15, 200, 229, 124, 215, 244, 93, 46, 186, 171, 201, 109, 140,
+  230, 62, 216, 245, 31, 125, 78, 156, 231, 47, 187, 202, 217, 94, 246, 141,
+  63, 232, 172, 110, 247, 157, 79, 218, 203, 126, 233, 188, 248, 95, 173, 142,
+  219, 111, 249, 234, 158, 127, 189, 204, 250, 235, 143, 174, 220, 205, 159,
+  251,
+  190, 221, 175, 236, 237, 191, 206, 252, 222, 253, 207, 238, 223, 254, 239,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_16x16[256]) = {
+  0, 16, 32, 48, 1, 64, 17, 80, 33, 96, 49, 2, 65, 112, 18, 81,
+  34, 128, 50, 97, 3, 66, 144, 19, 113, 35, 82, 160, 98, 51, 129, 4,
+  67, 176, 20, 114, 145, 83, 36, 99, 130, 52, 192, 5, 161, 68, 115, 21,
+  146, 84, 208, 177, 37, 131, 100, 53, 162, 224, 69, 6, 116, 193, 147, 85,
+  22, 240, 132, 38, 178, 101, 163, 54, 209, 117, 70, 7, 148, 194, 86, 179,
+  225, 23, 133, 39, 164, 8, 102, 210, 241, 55, 195, 118, 149, 71, 180, 24,
+  87, 226, 134, 165, 211, 40, 103, 56, 72, 150, 196, 242, 119, 9, 181, 227,
+  88, 166, 25, 135, 41, 104, 212, 57, 151, 197, 120, 73, 243, 182, 136, 167,
+  213, 89, 10, 228, 105, 152, 198, 26, 42, 121, 183, 244, 168, 58, 137, 229,
+  74, 214, 90, 153, 199, 184, 11, 106, 245, 27, 122, 230, 169, 43, 215, 59,
+  200, 138, 185, 246, 75, 12, 91, 154, 216, 231, 107, 28, 44, 201, 123, 170,
+  60, 247, 232, 76, 139, 13, 92, 217, 186, 248, 155, 108, 29, 124, 45, 202,
+  233, 171, 61, 14, 77, 140, 15, 249, 93, 30, 187, 156, 218, 46, 109, 125,
+  62, 172, 78, 203, 31, 141, 234, 94, 47, 188, 63, 157, 110, 250, 219, 79,
+  126, 204, 173, 142, 95, 189, 111, 235, 158, 220, 251, 127, 174, 143, 205,
+  236,
+  159, 190, 221, 252, 175, 206, 237, 191, 253, 222, 238, 207, 254, 223, 239,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_16x16[256]) = {
+  0, 1, 2, 16, 3, 17, 4, 18, 32, 5, 33, 19, 6, 34, 48, 20,
+  49, 7, 35, 21, 50, 64, 8, 36, 65, 22, 51, 37, 80, 9, 66, 52,
+  23, 38, 81, 67, 10, 53, 24, 82, 68, 96, 39, 11, 54, 83, 97, 69,
+  25, 98, 84, 40, 112, 55, 12, 70, 99, 113, 85, 26, 41, 56, 114, 100,
+  13, 71, 128, 86, 27, 115, 101, 129, 42, 57, 72, 116, 14, 87, 130, 102,
+  144, 73, 131, 117, 28, 58, 15, 88, 43, 145, 103, 132, 146, 118, 74, 160,
+  89, 133, 104, 29, 59, 147, 119, 44, 161, 148, 90, 105, 134, 162, 120, 176,
+  75, 135, 149, 30, 60, 163, 177, 45, 121, 91, 106, 164, 178, 150, 192, 136,
+  165, 179, 31, 151, 193, 76, 122, 61, 137, 194, 107, 152, 180, 208, 46, 166,
+  167, 195, 92, 181, 138, 209, 123, 153, 224, 196, 77, 168, 210, 182, 240, 108,
+  197, 62, 154, 225, 183, 169, 211, 47, 139, 93, 184, 226, 212, 241, 198, 170,
+  124, 155, 199, 78, 213, 185, 109, 227, 200, 63, 228, 242, 140, 214, 171, 186,
+  156, 229, 243, 125, 94, 201, 244, 215, 216, 230, 141, 187, 202, 79, 172, 110,
+  157, 245, 217, 231, 95, 246, 232, 126, 203, 247, 233, 173, 218, 142, 111,
+  158,
+  188, 248, 127, 234, 219, 249, 189, 204, 143, 174, 159, 250, 235, 205, 220,
+  175,
+  190, 251, 221, 191, 206, 236, 207, 237, 252, 222, 253, 223, 238, 239, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_32x32[1024]) = {
+  0, 32, 1, 64, 33, 2, 96, 65, 34, 128, 3, 97, 66, 160,
+  129, 35, 98, 4, 67, 130, 161, 192, 36, 99, 224, 5, 162, 193,
+  68, 131, 37, 100,
+  225, 194, 256, 163, 69, 132, 6, 226, 257, 288, 195, 101, 164, 38,
+  258, 7, 227, 289, 133, 320, 70, 196, 165, 290, 259, 228, 39, 321,
+  102, 352, 8, 197,
+  71, 134, 322, 291, 260, 353, 384, 229, 166, 103, 40, 354, 323, 292,
+  135, 385, 198, 261, 72, 9, 416, 167, 386, 355, 230, 324, 104, 293,
+  41, 417, 199, 136,
+  262, 387, 448, 325, 356, 10, 73, 418, 231, 168, 449, 294, 388, 105,
+  419, 263, 42, 200, 357, 450, 137, 480, 74, 326, 232, 11, 389, 169,
+  295, 420, 106, 451,
+  481, 358, 264, 327, 201, 43, 138, 512, 482, 390, 296, 233, 170, 421,
+  75, 452, 359, 12, 513, 265, 483, 328, 107, 202, 514, 544, 422, 391,
+  453, 139, 44, 234,
+  484, 297, 360, 171, 76, 515, 545, 266, 329, 454, 13, 423, 203, 108,
+  546, 485, 576, 298, 235, 140, 361, 330, 172, 547, 45, 455, 267, 577,
+  486, 77, 204, 362,
+  608, 14, 299, 578, 109, 236, 487, 609, 331, 141, 579, 46, 15, 173,
+  610, 363, 78, 205, 16, 110, 237, 611, 142, 47, 174, 79, 206, 17,
+  111, 238, 48, 143,
+  80, 175, 112, 207, 49, 18, 239, 81, 113, 19, 50, 82, 114, 51,
+  83, 115, 640, 516, 392, 268, 144, 20, 672, 641, 548, 517, 424,
+  393, 300, 269, 176, 145,
+  52, 21, 704, 673, 642, 580, 549, 518, 456, 425, 394, 332, 301,
+  270, 208, 177, 146, 84, 53, 22, 736, 705, 674, 643, 612, 581,
+  550, 519, 488, 457, 426, 395,
+  364, 333, 302, 271, 240, 209, 178, 147, 116, 85, 54, 23, 737,
+  706, 675, 613, 582, 551, 489, 458, 427, 365, 334, 303, 241,
+  210, 179, 117, 86, 55, 738, 707,
+  614, 583, 490, 459, 366, 335, 242, 211, 118, 87, 739, 615, 491,
+  367, 243, 119, 768, 644, 520, 396, 272, 148, 24, 800, 769, 676,
+  645, 552, 521, 428, 397, 304,
+  273, 180, 149, 56, 25, 832, 801, 770, 708, 677, 646, 584, 553,
+  522, 460, 429, 398, 336, 305, 274, 212, 181, 150, 88, 57, 26,
+  864, 833, 802, 771, 740, 709,
+  678, 647, 616, 585, 554, 523, 492, 461, 430, 399, 368, 337, 306,
+  275, 244, 213, 182, 151, 120, 89, 58, 27, 865, 834, 803, 741,
+  710, 679, 617, 586, 555, 493,
+  462, 431, 369, 338, 307, 245, 214, 183, 121, 90, 59, 866, 835,
+  742, 711, 618, 587, 494, 463, 370, 339, 246, 215, 122, 91, 867,
+  743, 619, 495, 371, 247, 123,
+  896, 772, 648, 524, 400, 276, 152, 28, 928, 897, 804, 773, 680,
+  649, 556, 525, 432, 401, 308, 277, 184, 153, 60, 29, 960, 929,
+  898, 836, 805, 774, 712, 681,
+  650, 588, 557, 526, 464, 433, 402, 340, 309, 278, 216, 185, 154,
+  92, 61, 30, 992, 961, 930, 899, 868, 837, 806, 775, 744, 713, 682,
+  651, 620, 589, 558, 527,
+  496, 465, 434, 403, 372, 341, 310, 279, 248, 217, 186, 155, 124,
+  93, 62, 31, 993, 962, 931, 869, 838, 807, 745, 714, 683, 621, 590,
+  559, 497, 466, 435, 373,
+  342, 311, 249, 218, 187, 125, 94, 63, 994, 963, 870, 839, 746, 715,
+  622, 591, 498, 467, 374, 343, 250, 219, 126, 95, 995, 871, 747, 623,
+  499, 375, 251, 127,
+  900, 776, 652, 528, 404, 280, 156, 932, 901, 808, 777, 684, 653, 560,
+  529, 436, 405, 312, 281, 188, 157, 964, 933, 902, 840, 809, 778, 716,
+  685, 654, 592, 561,
+  530, 468, 437, 406, 344, 313, 282, 220, 189, 158, 996, 965, 934, 903,
+  872, 841, 810, 779, 748, 717, 686, 655, 624, 593, 562, 531, 500, 469,
+  438, 407, 376, 345,
+  314, 283, 252, 221, 190, 159, 997, 966, 935, 873, 842, 811, 749, 718,
+  687, 625, 594, 563, 501, 470, 439, 377, 346, 315, 253, 222, 191, 998,
+  967, 874, 843, 750,
+  719, 626, 595, 502, 471, 378, 347, 254, 223, 999, 875, 751, 627, 503,
+  379, 255, 904, 780, 656, 532, 408, 284, 936, 905, 812, 781, 688, 657,
+  564, 533, 440, 409,
+  316, 285, 968, 937, 906, 844, 813, 782, 720, 689, 658, 596, 565, 534,
+  472, 441, 410, 348, 317, 286, 1000, 969, 938, 907, 876, 845, 814, 783,
+  752, 721, 690, 659,
+  628, 597, 566, 535, 504, 473, 442, 411, 380, 349, 318, 287, 1001, 970,
+  939, 877, 846, 815, 753, 722, 691, 629, 598, 567, 505, 474, 443, 381,
+  350, 319, 1002, 971,
+  878, 847, 754, 723, 630, 599, 506, 475, 382, 351, 1003, 879, 755, 631,
+  507, 383, 908, 784, 660, 536, 412, 940, 909, 816, 785, 692, 661, 568,
+  537, 444, 413, 972,
+  941, 910, 848, 817, 786, 724, 693, 662, 600, 569, 538, 476, 445, 414,
+  1004, 973, 942, 911, 880, 849, 818, 787, 756, 725, 694, 663, 632, 601,
+  570, 539, 508, 477,
+  446, 415, 1005, 974, 943, 881, 850, 819, 757, 726, 695, 633, 602, 571,
+  509, 478, 447, 1006, 975, 882, 851, 758, 727, 634, 603, 510, 479,
+  1007, 883, 759, 635, 511,
+  912, 788, 664, 540, 944, 913, 820, 789, 696, 665, 572, 541, 976, 945,
+  914, 852, 821, 790, 728, 697, 666, 604, 573, 542, 1008, 977, 946, 915,
+  884, 853, 822, 791,
+  760, 729, 698, 667, 636, 605, 574, 543, 1009, 978, 947, 885, 854, 823,
+  761, 730, 699, 637, 606, 575, 1010, 979, 886, 855, 762, 731, 638, 607,
+  1011, 887, 763, 639,
+  916, 792, 668, 948, 917, 824, 793, 700, 669, 980, 949, 918, 856, 825,
+  794, 732, 701, 670, 1012, 981, 950, 919, 888, 857, 826, 795, 764, 733,
+  702, 671, 1013, 982,
+  951, 889, 858, 827, 765, 734, 703, 1014, 983, 890, 859, 766, 735, 1015,
+  891, 767, 920, 796, 952, 921, 828, 797, 984, 953, 922, 860, 829, 798,
+  1016, 985, 954, 923,
+  892, 861, 830, 799, 1017, 986, 955, 893, 862, 831, 1018, 987, 894, 863,
+  1019, 895, 924, 956, 925, 988, 957, 926, 1020, 989, 958, 927, 1021,
+  990, 959, 1022, 991, 1023,
+};
+
+// Neighborhood 5-tuples for various scans and blocksizes,
+// in {top, left, topleft, topright, bottomleft} order
+// for each position in raster scan order.
+// -1 indicates the neighbor does not exist.
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 1, 4, 4, 4, 1, 1, 8, 8, 5, 8, 2, 2, 2, 5, 9, 12, 6, 9,
+  3, 6, 10, 13, 7, 10, 11, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 4, 4, 0, 0, 8, 8, 1, 1, 5, 5, 1, 1, 9, 9, 2, 2, 6, 6, 2, 2, 3,
+  3, 10, 10, 7, 7, 11, 11, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 1, 1, 4, 4, 2, 2, 5, 5, 4, 4, 8, 8, 6, 6, 8, 8, 9, 9, 12,
+  12, 10, 10, 13, 13, 14, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 8, 8, 0, 0, 16, 16, 1, 1, 24, 24, 9, 9, 1, 1, 32, 32, 17, 17, 2,
+  2, 25, 25, 10, 10, 40, 40, 2, 2, 18, 18, 33, 33, 3, 3, 48, 48, 11, 11, 26,
+  26, 3, 3, 41, 41, 19, 19, 34, 34, 4, 4, 27, 27, 12, 12, 49, 49, 42, 42, 20,
+  20, 4, 4, 35, 35, 5, 5, 28, 28, 50, 50, 43, 43, 13, 13, 36, 36, 5, 5, 21, 21,
+  51, 51, 29, 29, 6, 6, 44, 44, 14, 14, 6, 6, 37, 37, 52, 52, 22, 22, 7, 7, 30,
+  30, 45, 45, 15, 15, 38, 38, 23, 23, 53, 53, 31, 31, 46, 46, 39, 39, 54, 54,
+  47, 47, 55, 55, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 1, 1, 0, 0, 8, 8, 2, 2, 8, 8, 9, 9, 3, 3, 16, 16, 10, 10, 16, 16,
+  4, 4, 17, 17, 24, 24, 11, 11, 18, 18, 25, 25, 24, 24, 5, 5, 12, 12, 19, 19,
+  32, 32, 26, 26, 6, 6, 33, 33, 32, 32, 20, 20, 27, 27, 40, 40, 13, 13, 34, 34,
+  40, 40, 41, 41, 28, 28, 35, 35, 48, 48, 21, 21, 42, 42, 14, 14, 48, 48, 36,
+  36, 49, 49, 43, 43, 29, 29, 56, 56, 22, 22, 50, 50, 57, 57, 44, 44, 37, 37,
+  51, 51, 30, 30, 58, 58, 52, 52, 45, 45, 59, 59, 38, 38, 60, 60, 46, 46, 53,
+  53, 54, 54, 61, 61, 62, 62, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 8, 8, 1, 8, 1, 1, 9, 16, 16, 16, 2, 9, 2, 2, 10, 17, 17,
+  24, 24, 24, 3, 10, 3, 3, 18, 25, 25, 32, 11, 18, 32, 32, 4, 11, 26, 33, 19,
+  26, 4, 4, 33, 40, 12, 19, 40, 40, 5, 12, 27, 34, 34, 41, 20, 27, 13, 20, 5,
+  5, 41, 48, 48, 48, 28, 35, 35, 42, 21, 28, 6, 6, 6, 13, 42, 49, 49, 56, 36,
+  43, 14, 21, 29, 36, 7, 14, 43, 50, 50, 57, 22, 29, 37, 44, 15, 22, 44, 51,
+  51, 58, 30, 37, 23, 30, 52, 59, 45, 52, 38, 45, 31, 38, 53, 60, 46, 53, 39,
+  46, 54, 61, 47, 54, 55, 62, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 16, 16, 32, 32, 0, 0, 48, 48, 1, 1, 64, 64,
+  17, 17, 80, 80, 33, 33, 1, 1, 49, 49, 96, 96, 2, 2, 65, 65,
+  18, 18, 112, 112, 34, 34, 81, 81, 2, 2, 50, 50, 128, 128, 3, 3,
+  97, 97, 19, 19, 66, 66, 144, 144, 82, 82, 35, 35, 113, 113, 3, 3,
+  51, 51, 160, 160, 4, 4, 98, 98, 129, 129, 67, 67, 20, 20, 83, 83,
+  114, 114, 36, 36, 176, 176, 4, 4, 145, 145, 52, 52, 99, 99, 5, 5,
+  130, 130, 68, 68, 192, 192, 161, 161, 21, 21, 115, 115, 84, 84, 37, 37,
+  146, 146, 208, 208, 53, 53, 5, 5, 100, 100, 177, 177, 131, 131, 69, 69,
+  6, 6, 224, 224, 116, 116, 22, 22, 162, 162, 85, 85, 147, 147, 38, 38,
+  193, 193, 101, 101, 54, 54, 6, 6, 132, 132, 178, 178, 70, 70, 163, 163,
+  209, 209, 7, 7, 117, 117, 23, 23, 148, 148, 7, 7, 86, 86, 194, 194,
+  225, 225, 39, 39, 179, 179, 102, 102, 133, 133, 55, 55, 164, 164, 8, 8,
+  71, 71, 210, 210, 118, 118, 149, 149, 195, 195, 24, 24, 87, 87, 40, 40,
+  56, 56, 134, 134, 180, 180, 226, 226, 103, 103, 8, 8, 165, 165, 211, 211,
+  72, 72, 150, 150, 9, 9, 119, 119, 25, 25, 88, 88, 196, 196, 41, 41,
+  135, 135, 181, 181, 104, 104, 57, 57, 227, 227, 166, 166, 120, 120, 151, 151,
+  197, 197, 73, 73, 9, 9, 212, 212, 89, 89, 136, 136, 182, 182, 10, 10,
+  26, 26, 105, 105, 167, 167, 228, 228, 152, 152, 42, 42, 121, 121, 213, 213,
+  58, 58, 198, 198, 74, 74, 137, 137, 183, 183, 168, 168, 10, 10, 90, 90,
+  229, 229, 11, 11, 106, 106, 214, 214, 153, 153, 27, 27, 199, 199, 43, 43,
+  184, 184, 122, 122, 169, 169, 230, 230, 59, 59, 11, 11, 75, 75, 138, 138,
+  200, 200, 215, 215, 91, 91, 12, 12, 28, 28, 185, 185, 107, 107, 154, 154,
+  44, 44, 231, 231, 216, 216, 60, 60, 123, 123, 12, 12, 76, 76, 201, 201,
+  170, 170, 232, 232, 139, 139, 92, 92, 13, 13, 108, 108, 29, 29, 186, 186,
+  217, 217, 155, 155, 45, 45, 13, 13, 61, 61, 124, 124, 14, 14, 233, 233,
+  77, 77, 14, 14, 171, 171, 140, 140, 202, 202, 30, 30, 93, 93, 109, 109,
+  46, 46, 156, 156, 62, 62, 187, 187, 15, 15, 125, 125, 218, 218, 78, 78,
+  31, 31, 172, 172, 47, 47, 141, 141, 94, 94, 234, 234, 203, 203, 63, 63,
+  110, 110, 188, 188, 157, 157, 126, 126, 79, 79, 173, 173, 95, 95, 219, 219,
+  142, 142, 204, 204, 235, 235, 111, 111, 158, 158, 127, 127, 189, 189, 220,
+  220, 143, 143, 174, 174, 205, 205, 236, 236, 159, 159, 190, 190, 221, 221,
+  175, 175, 237, 237, 206, 206, 222, 222, 191, 191, 238, 238, 207, 207, 223,
+  223, 239, 239, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 1, 1, 0, 0, 2, 2, 16, 16, 3, 3, 17, 17,
+  16, 16, 4, 4, 32, 32, 18, 18, 5, 5, 33, 33, 32, 32, 19, 19,
+  48, 48, 6, 6, 34, 34, 20, 20, 49, 49, 48, 48, 7, 7, 35, 35,
+  64, 64, 21, 21, 50, 50, 36, 36, 64, 64, 8, 8, 65, 65, 51, 51,
+  22, 22, 37, 37, 80, 80, 66, 66, 9, 9, 52, 52, 23, 23, 81, 81,
+  67, 67, 80, 80, 38, 38, 10, 10, 53, 53, 82, 82, 96, 96, 68, 68,
+  24, 24, 97, 97, 83, 83, 39, 39, 96, 96, 54, 54, 11, 11, 69, 69,
+  98, 98, 112, 112, 84, 84, 25, 25, 40, 40, 55, 55, 113, 113, 99, 99,
+  12, 12, 70, 70, 112, 112, 85, 85, 26, 26, 114, 114, 100, 100, 128, 128,
+  41, 41, 56, 56, 71, 71, 115, 115, 13, 13, 86, 86, 129, 129, 101, 101,
+  128, 128, 72, 72, 130, 130, 116, 116, 27, 27, 57, 57, 14, 14, 87, 87,
+  42, 42, 144, 144, 102, 102, 131, 131, 145, 145, 117, 117, 73, 73, 144, 144,
+  88, 88, 132, 132, 103, 103, 28, 28, 58, 58, 146, 146, 118, 118, 43, 43,
+  160, 160, 147, 147, 89, 89, 104, 104, 133, 133, 161, 161, 119, 119, 160, 160,
+  74, 74, 134, 134, 148, 148, 29, 29, 59, 59, 162, 162, 176, 176, 44, 44,
+  120, 120, 90, 90, 105, 105, 163, 163, 177, 177, 149, 149, 176, 176, 135, 135,
+  164, 164, 178, 178, 30, 30, 150, 150, 192, 192, 75, 75, 121, 121, 60, 60,
+  136, 136, 193, 193, 106, 106, 151, 151, 179, 179, 192, 192, 45, 45, 165, 165,
+  166, 166, 194, 194, 91, 91, 180, 180, 137, 137, 208, 208, 122, 122, 152, 152,
+  208, 208, 195, 195, 76, 76, 167, 167, 209, 209, 181, 181, 224, 224, 107, 107,
+  196, 196, 61, 61, 153, 153, 224, 224, 182, 182, 168, 168, 210, 210, 46, 46,
+  138, 138, 92, 92, 183, 183, 225, 225, 211, 211, 240, 240, 197, 197, 169, 169,
+  123, 123, 154, 154, 198, 198, 77, 77, 212, 212, 184, 184, 108, 108, 226, 226,
+  199, 199, 62, 62, 227, 227, 241, 241, 139, 139, 213, 213, 170, 170, 185, 185,
+  155, 155, 228, 228, 242, 242, 124, 124, 93, 93, 200, 200, 243, 243, 214, 214,
+  215, 215, 229, 229, 140, 140, 186, 186, 201, 201, 78, 78, 171, 171, 109, 109,
+  156, 156, 244, 244, 216, 216, 230, 230, 94, 94, 245, 245, 231, 231, 125, 125,
+  202, 202, 246, 246, 232, 232, 172, 172, 217, 217, 141, 141, 110, 110, 157,
+  157, 187, 187, 247, 247, 126, 126, 233, 233, 218, 218, 248, 248, 188, 188,
+  203, 203, 142, 142, 173, 173, 158, 158, 249, 249, 234, 234, 204, 204, 219,
+  219, 174, 174, 189, 189, 250, 250, 220, 220, 190, 190, 205, 205, 235, 235,
+  206, 206, 236, 236, 251, 251, 221, 221, 252, 252, 222, 222, 237, 237, 238,
+  238, 253, 253, 254, 254, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 16, 16, 1, 16, 1, 1, 32, 32, 17, 32,
+  2, 17, 2, 2, 48, 48, 18, 33, 33, 48, 3, 18, 49, 64, 64, 64,
+  34, 49, 3, 3, 19, 34, 50, 65, 4, 19, 65, 80, 80, 80, 35, 50,
+  4, 4, 20, 35, 66, 81, 81, 96, 51, 66, 96, 96, 5, 20, 36, 51,
+  82, 97, 21, 36, 67, 82, 97, 112, 5, 5, 52, 67, 112, 112, 37, 52,
+  6, 21, 83, 98, 98, 113, 68, 83, 6, 6, 113, 128, 22, 37, 53, 68,
+  84, 99, 99, 114, 128, 128, 114, 129, 69, 84, 38, 53, 7, 22, 7, 7,
+  129, 144, 23, 38, 54, 69, 100, 115, 85, 100, 115, 130, 144, 144, 130, 145,
+  39, 54, 70, 85, 8, 23, 55, 70, 116, 131, 101, 116, 145, 160, 24, 39,
+  8, 8, 86, 101, 131, 146, 160, 160, 146, 161, 71, 86, 40, 55, 9, 24,
+  117, 132, 102, 117, 161, 176, 132, 147, 56, 71, 87, 102, 25, 40, 147, 162,
+  9, 9, 176, 176, 162, 177, 72, 87, 41, 56, 118, 133, 133, 148, 103, 118,
+  10, 25, 148, 163, 57, 72, 88, 103, 177, 192, 26, 41, 163, 178, 192, 192,
+  10, 10, 119, 134, 73, 88, 149, 164, 104, 119, 134, 149, 42, 57, 178, 193,
+  164, 179, 11, 26, 58, 73, 193, 208, 89, 104, 135, 150, 120, 135, 27, 42,
+  74, 89, 208, 208, 150, 165, 179, 194, 165, 180, 105, 120, 194, 209, 43, 58,
+  11, 11, 136, 151, 90, 105, 151, 166, 180, 195, 59, 74, 121, 136, 209, 224,
+  195, 210, 224, 224, 166, 181, 106, 121, 75, 90, 12, 27, 181, 196, 12, 12,
+  210, 225, 152, 167, 167, 182, 137, 152, 28, 43, 196, 211, 122, 137, 91, 106,
+  225, 240, 44, 59, 13, 28, 107, 122, 182, 197, 168, 183, 211, 226, 153, 168,
+  226, 241, 60, 75, 197, 212, 138, 153, 29, 44, 76, 91, 13, 13, 183, 198,
+  123, 138, 45, 60, 212, 227, 198, 213, 154, 169, 169, 184, 227, 242, 92, 107,
+  61, 76, 139, 154, 14, 29, 14, 14, 184, 199, 213, 228, 108, 123, 199, 214,
+  228, 243, 77, 92, 30, 45, 170, 185, 155, 170, 185, 200, 93, 108, 124, 139,
+  214, 229, 46, 61, 200, 215, 229, 244, 15, 30, 109, 124, 62, 77, 140, 155,
+  215, 230, 31, 46, 171, 186, 186, 201, 201, 216, 78, 93, 230, 245, 125, 140,
+  47, 62, 216, 231, 156, 171, 94, 109, 231, 246, 141, 156, 63, 78, 202, 217,
+  187, 202, 110, 125, 217, 232, 172, 187, 232, 247, 79, 94, 157, 172, 126, 141,
+  203, 218, 95, 110, 233, 248, 218, 233, 142, 157, 111, 126, 173, 188, 188, 203,
+  234, 249, 219, 234, 127, 142, 158, 173, 204, 219, 189, 204, 143, 158, 235,
+  250, 174, 189, 205, 220, 159, 174, 220, 235, 221, 236, 175, 190, 190, 205,
+  236, 251, 206, 221, 237, 252, 191, 206, 222, 237, 207, 222, 238, 253, 223,
+  238, 239, 254, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 0, 32, 32, 1, 32, 1, 1, 64, 64, 33, 64,
+  2, 33, 96, 96, 2, 2, 65, 96, 34, 65, 128, 128, 97, 128, 3, 34,
+  66, 97, 3, 3, 35, 66, 98, 129, 129, 160, 160, 160, 4, 35, 67, 98,
+  192, 192, 4, 4, 130, 161, 161, 192, 36, 67, 99, 130, 5, 36, 68, 99,
+  193, 224, 162, 193, 224, 224, 131, 162, 37, 68, 100, 131, 5, 5, 194, 225,
+  225, 256, 256, 256, 163, 194, 69, 100, 132, 163, 6, 37, 226, 257, 6, 6,
+  195, 226, 257, 288, 101, 132, 288, 288, 38, 69, 164, 195, 133, 164, 258, 289,
+  227, 258, 196, 227, 7, 38, 289, 320, 70, 101, 320, 320, 7, 7, 165, 196,
+  39, 70, 102, 133, 290, 321, 259, 290, 228, 259, 321, 352, 352, 352, 197, 228,
+  134, 165, 71, 102, 8, 39, 322, 353, 291, 322, 260, 291, 103, 134, 353, 384,
+  166, 197, 229, 260, 40, 71, 8, 8, 384, 384, 135, 166, 354, 385, 323, 354,
+  198, 229, 292, 323, 72, 103, 261, 292, 9, 40, 385, 416, 167, 198, 104, 135,
+  230, 261, 355, 386, 416, 416, 293, 324, 324, 355, 9, 9, 41, 72, 386, 417,
+  199, 230, 136, 167, 417, 448, 262, 293, 356, 387, 73, 104, 387, 418, 231, 262,
+  10, 41, 168, 199, 325, 356, 418, 449, 105, 136, 448, 448, 42, 73, 294, 325,
+  200, 231, 10, 10, 357, 388, 137, 168, 263, 294, 388, 419, 74, 105, 419, 450,
+  449, 480, 326, 357, 232, 263, 295, 326, 169, 200, 11, 42, 106, 137, 480, 480,
+  450, 481, 358, 389, 264, 295, 201, 232, 138, 169, 389, 420, 43, 74, 420, 451,
+  327, 358, 11, 11, 481, 512, 233, 264, 451, 482, 296, 327, 75, 106, 170, 201,
+  482, 513, 512, 512, 390, 421, 359, 390, 421, 452, 107, 138, 12, 43, 202, 233,
+  452, 483, 265, 296, 328, 359, 139, 170, 44, 75, 483, 514, 513, 544, 234, 265,
+  297, 328, 422, 453, 12, 12, 391, 422, 171, 202, 76, 107, 514, 545, 453, 484,
+  544, 544, 266, 297, 203, 234, 108, 139, 329, 360, 298, 329, 140, 171, 515,
+  546, 13, 44, 423, 454, 235, 266, 545, 576, 454, 485, 45, 76, 172, 203, 330,
+  361, 576, 576, 13, 13, 267, 298, 546, 577, 77, 108, 204, 235, 455, 486, 577,
+  608, 299, 330, 109, 140, 547, 578, 14, 45, 14, 14, 141, 172, 578, 609, 331,
+  362, 46, 77, 173, 204, 15, 15, 78, 109, 205, 236, 579, 610, 110, 141, 15, 46,
+  142, 173, 47, 78, 174, 205, 16, 16, 79, 110, 206, 237, 16, 47, 111, 142,
+  48, 79, 143, 174, 80, 111, 175, 206, 17, 48, 17, 17, 207, 238, 49, 80,
+  81, 112, 18, 18, 18, 49, 50, 81, 82, 113, 19, 50, 51, 82, 83, 114, 608, 608,
+  484, 515, 360, 391, 236, 267, 112, 143, 19, 19, 640, 640, 609, 640, 516, 547,
+  485, 516, 392, 423, 361, 392, 268, 299, 237, 268, 144, 175, 113, 144, 20, 51,
+  20, 20, 672, 672, 641, 672, 610, 641, 548, 579, 517, 548, 486, 517, 424, 455,
+  393, 424, 362, 393, 300, 331, 269, 300, 238, 269, 176, 207, 145, 176, 114,
+  145, 52, 83, 21, 52, 21, 21, 704, 704, 673, 704, 642, 673, 611, 642, 580,
+  611, 549, 580, 518, 549, 487, 518, 456, 487, 425, 456, 394, 425, 363, 394,
+  332, 363, 301, 332, 270, 301, 239, 270, 208, 239, 177, 208, 146, 177, 115,
+  146, 84, 115, 53, 84, 22, 53, 22, 22, 705, 736, 674, 705, 643, 674, 581, 612,
+  550, 581, 519, 550, 457, 488, 426, 457, 395, 426, 333, 364, 302, 333, 271,
+  302, 209, 240, 178, 209, 147, 178, 85, 116, 54, 85, 23, 54, 706, 737, 675,
+  706, 582, 613, 551, 582, 458, 489, 427, 458, 334, 365, 303, 334, 210, 241,
+  179, 210, 86, 117, 55, 86, 707, 738, 583, 614, 459, 490, 335, 366, 211, 242,
+  87, 118, 736, 736, 612, 643, 488, 519, 364, 395, 240, 271, 116, 147, 23, 23,
+  768, 768, 737, 768, 644, 675, 613, 644, 520, 551, 489, 520, 396, 427, 365,
+  396, 272, 303, 241, 272, 148, 179, 117, 148, 24, 55, 24, 24, 800, 800, 769,
+  800, 738, 769, 676, 707, 645, 676, 614, 645, 552, 583, 521, 552, 490, 521,
+  428, 459, 397, 428, 366, 397, 304, 335, 273, 304, 242, 273, 180, 211, 149,
+  180, 118, 149, 56, 87, 25, 56, 25, 25, 832, 832, 801, 832, 770, 801, 739,
+  770, 708, 739, 677, 708, 646, 677, 615, 646, 584, 615, 553, 584, 522, 553,
+  491, 522, 460, 491, 429, 460, 398, 429, 367, 398, 336, 367, 305, 336, 274,
+  305, 243, 274, 212, 243, 181, 212, 150, 181, 119, 150, 88, 119, 57, 88, 26,
+  57, 26, 26, 833, 864, 802, 833, 771, 802, 709, 740, 678, 709, 647, 678, 585,
+  616, 554, 585, 523, 554, 461, 492, 430, 461, 399, 430, 337, 368, 306, 337,
+  275, 306, 213, 244, 182, 213, 151, 182, 89, 120, 58, 89, 27, 58, 834, 865,
+  803, 834, 710, 741, 679, 710, 586, 617, 555, 586, 462, 493, 431, 462, 338,
+  369, 307, 338, 214, 245, 183, 214, 90, 121, 59, 90, 835, 866, 711, 742, 587,
+  618, 463, 494, 339, 370, 215, 246, 91, 122, 864, 864, 740, 771, 616, 647,
+  492, 523, 368, 399, 244, 275, 120, 151, 27, 27, 896, 896, 865, 896, 772, 803,
+  741, 772, 648, 679, 617, 648, 524, 555, 493, 524, 400, 431, 369, 400, 276,
+  307, 245, 276, 152, 183, 121, 152, 28, 59, 28, 28, 928, 928, 897, 928, 866,
+  897, 804, 835, 773, 804, 742, 773, 680, 711, 649, 680, 618, 649, 556, 587,
+  525, 556, 494, 525, 432, 463, 401, 432, 370, 401, 308, 339, 277, 308, 246,
+  277, 184, 215, 153, 184, 122, 153, 60, 91, 29, 60, 29, 29, 960, 960, 929,
+  960, 898, 929, 867, 898, 836, 867, 805, 836, 774, 805, 743, 774, 712, 743,
+  681, 712, 650, 681, 619, 650, 588, 619, 557, 588, 526, 557, 495, 526, 464,
+  495, 433, 464, 402, 433, 371, 402, 340, 371, 309, 340, 278, 309, 247, 278,
+  216, 247, 185, 216, 154, 185, 123, 154, 92, 123, 61, 92, 30, 61, 30, 30,
+  961, 992, 930, 961, 899, 930, 837, 868, 806, 837, 775, 806, 713, 744, 682,
+  713, 651, 682, 589, 620, 558, 589, 527, 558, 465, 496, 434, 465, 403, 434,
+  341, 372, 310, 341, 279, 310, 217, 248, 186, 217, 155, 186, 93, 124, 62, 93,
+  31, 62, 962, 993, 931, 962, 838, 869, 807, 838, 714, 745, 683, 714, 590, 621,
+  559, 590, 466, 497, 435, 466, 342, 373, 311, 342, 218, 249, 187, 218, 94,
+  125, 63, 94, 963, 994, 839, 870, 715, 746, 591, 622, 467, 498, 343, 374, 219,
+  250, 95, 126, 868, 899, 744, 775, 620, 651, 496, 527, 372, 403, 248, 279,
+  124, 155, 900, 931, 869, 900, 776, 807, 745, 776, 652, 683, 621, 652, 528,
+  559, 497, 528, 404, 435, 373, 404, 280, 311, 249, 280, 156, 187, 125, 156,
+  932, 963, 901, 932, 870, 901, 808, 839, 777, 808, 746, 777, 684, 715, 653,
+  684, 622, 653, 560, 591, 529, 560, 498, 529, 436, 467, 405, 436, 374, 405,
+  312, 343, 281, 312, 250, 281, 188, 219, 157, 188, 126, 157, 964, 995, 933,
+  964, 902, 933, 871, 902, 840, 871, 809, 840, 778, 809, 747, 778, 716, 747,
+  685, 716, 654, 685, 623, 654, 592, 623, 561, 592, 530, 561, 499, 530, 468,
+  499, 437, 468, 406, 437, 375, 406, 344, 375, 313, 344, 282, 313, 251, 282,
+  220, 251, 189, 220, 158, 189, 127, 158, 965, 996, 934, 965, 903, 934, 841,
+  872, 810, 841, 779, 810, 717, 748, 686, 717, 655, 686, 593, 624, 562, 593,
+  531, 562, 469, 500, 438, 469, 407, 438, 345, 376, 314, 345, 283, 314, 221,
+  252, 190, 221, 159, 190, 966, 997, 935, 966, 842, 873, 811, 842, 718, 749,
+  687, 718, 594, 625, 563, 594, 470, 501, 439, 470, 346, 377, 315, 346, 222,
+  253, 191, 222, 967, 998, 843, 874, 719, 750, 595, 626, 471, 502, 347, 378,
+  223, 254, 872, 903, 748, 779, 624, 655, 500, 531, 376, 407, 252, 283, 904,
+  935, 873, 904, 780, 811, 749, 780, 656, 687, 625, 656, 532, 563, 501, 532,
+  408, 439, 377, 408, 284, 315, 253, 284, 936, 967, 905, 936, 874, 905, 812,
+  843, 781, 812, 750, 781, 688, 719, 657, 688, 626, 657, 564, 595, 533, 564,
+  502, 533, 440, 471, 409, 440, 378, 409, 316, 347, 285, 316, 254, 285, 968,
+  999, 937, 968, 906, 937, 875, 906, 844, 875, 813, 844, 782, 813, 751, 782,
+  720, 751, 689, 720, 658, 689, 627, 658, 596, 627, 565, 596, 534, 565, 503,
+  534, 472, 503, 441, 472, 410, 441, 379, 410, 348, 379, 317, 348, 286, 317,
+  255, 286, 969, 1000, 938, 969, 907, 938, 845, 876, 814, 845, 783, 814, 721,
+  752, 690, 721, 659, 690, 597, 628, 566, 597, 535, 566, 473, 504, 442, 473,
+  411, 442, 349, 380, 318, 349, 287, 318, 970, 1001, 939, 970, 846, 877, 815,
+  846, 722, 753, 691, 722, 598, 629, 567, 598, 474, 505, 443, 474, 350, 381,
+  319, 350, 971, 1002, 847, 878, 723, 754, 599, 630, 475, 506, 351, 382, 876,
+  907, 752, 783, 628, 659, 504, 535, 380, 411, 908, 939, 877, 908, 784, 815,
+  753, 784, 660, 691, 629, 660, 536, 567, 505, 536, 412, 443, 381, 412, 940,
+  971, 909, 940, 878, 909, 816, 847, 785, 816, 754, 785, 692, 723, 661, 692,
+  630, 661, 568, 599, 537, 568, 506, 537, 444, 475, 413, 444, 382, 413, 972,
+  1003, 941, 972, 910, 941, 879, 910, 848, 879, 817, 848, 786, 817, 755, 786,
+  724, 755, 693, 724, 662, 693, 631, 662, 600, 631, 569, 600, 538, 569, 507,
+  538, 476, 507, 445, 476, 414, 445, 383, 414, 973, 1004, 942, 973, 911, 942,
+  849, 880, 818, 849, 787, 818, 725, 756, 694, 725, 663, 694, 601, 632, 570,
+  601, 539, 570, 477, 508, 446, 477, 415, 446, 974, 1005, 943, 974, 850, 881,
+  819, 850, 726, 757, 695, 726, 602, 633, 571, 602, 478, 509, 447, 478, 975,
+  1006, 851, 882, 727, 758, 603, 634, 479, 510, 880, 911, 756, 787, 632, 663,
+  508, 539, 912, 943, 881, 912, 788, 819, 757, 788, 664, 695, 633, 664, 540,
+  571, 509, 540, 944, 975, 913, 944, 882, 913, 820, 851, 789, 820, 758, 789,
+  696, 727, 665, 696, 634, 665, 572, 603, 541, 572, 510, 541, 976, 1007, 945,
+  976, 914, 945, 883, 914, 852, 883, 821, 852, 790, 821, 759, 790, 728, 759,
+  697, 728, 666, 697, 635, 666, 604, 635, 573, 604, 542, 573, 511, 542, 977,
+  1008, 946, 977, 915, 946, 853, 884, 822, 853, 791, 822, 729, 760, 698, 729,
+  667, 698, 605, 636, 574, 605, 543, 574, 978, 1009, 947, 978, 854, 885, 823,
+  854, 730, 761, 699, 730, 606, 637, 575, 606, 979, 1010, 855, 886, 731, 762,
+  607, 638, 884, 915, 760, 791, 636, 667, 916, 947, 885, 916, 792, 823, 761,
+  792, 668, 699, 637, 668, 948, 979, 917, 948, 886, 917, 824, 855, 793, 824,
+  762, 793, 700, 731, 669, 700, 638, 669, 980, 1011, 949, 980, 918, 949, 887,
+  918, 856, 887, 825, 856, 794, 825, 763, 794, 732, 763, 701, 732, 670, 701,
+  639, 670, 981, 1012, 950, 981, 919, 950, 857, 888, 826, 857, 795, 826, 733,
+  764, 702, 733, 671, 702, 982, 1013, 951, 982, 858, 889, 827, 858, 734, 765,
+  703, 734, 983, 1014, 859, 890, 735, 766, 888, 919, 764, 795, 920, 951, 889,
+  920, 796, 827, 765, 796, 952, 983, 921, 952, 890, 921, 828, 859, 797, 828,
+  766, 797, 984, 1015, 953, 984, 922, 953, 891, 922, 860, 891, 829, 860, 798,
+  829, 767, 798, 985, 1016, 954, 985, 923, 954, 861, 892, 830, 861, 799, 830,
+  986, 1017, 955, 986, 862, 893, 831, 862, 987, 1018, 863, 894, 892, 923, 924,
+  955, 893, 924, 956, 987, 925, 956, 894, 925, 988, 1019, 957, 988, 926, 957,
+  895, 926, 989, 1020, 958, 989, 927, 958, 990, 1021, 959, 990, 991, 1022, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_4x4[16]) = {
+  0, 2, 5, 8, 1, 3, 9, 12, 4, 7, 11, 14, 6, 10, 13, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_4x4[16]) = {
+  0, 3, 7, 11, 1, 5, 9, 12, 2, 6, 10, 14, 4, 8, 13, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_4x4[16]) = {
+  0, 1, 3, 5, 2, 4, 6, 9, 7, 8, 11, 13, 10, 12, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_8x8[64]) = {
+  0, 3, 8, 15, 22, 32, 40, 47, 1, 5, 11, 18, 26, 34, 44, 51,
+  2, 7, 13, 20, 28, 38, 46, 54, 4, 10, 16, 24, 31, 41, 50, 56,
+  6, 12, 21, 27, 35, 43, 52, 58, 9, 17, 25, 33, 39, 48, 55, 60,
+  14, 23, 30, 37, 45, 53, 59, 62, 19, 29, 36, 42, 49, 57, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_8x8[64]) = {
+  0, 1, 2, 5, 8, 12, 19, 24, 3, 4, 7, 10, 15, 20, 30, 39,
+  6, 9, 13, 16, 21, 27, 37, 46, 11, 14, 17, 23, 28, 34, 44, 52,
+  18, 22, 25, 31, 35, 41, 50, 57, 26, 29, 33, 38, 43, 49, 55, 59,
+  32, 36, 42, 47, 51, 54, 60, 61, 40, 45, 48, 53, 56, 58, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_8x8[64]) = {
+  0, 2, 5, 9, 14, 22, 31, 37, 1, 4, 8, 13, 19, 26, 38, 44,
+  3, 6, 10, 17, 24, 30, 42, 49, 7, 11, 15, 21, 29, 36, 47, 53,
+  12, 16, 20, 27, 34, 43, 52, 57, 18, 23, 28, 35, 41, 48, 56, 60,
+  25, 32, 39, 45, 50, 55, 59, 62, 33, 40, 46, 51, 54, 58, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_col_iscan_16x16[256]) = {
+  0, 4, 11, 20, 31, 43, 59, 75, 85, 109, 130, 150, 165, 181, 195, 198,
+  1, 6, 14, 23, 34, 47, 64, 81, 95, 114, 135, 153, 171, 188, 201, 212,
+  2, 8, 16, 25, 38, 52, 67, 83, 101, 116, 136, 157, 172, 190, 205, 216,
+  3, 10, 18, 29, 41, 55, 71, 89, 103, 119, 141, 159, 176, 194, 208, 218,
+  5, 12, 21, 32, 45, 58, 74, 93, 104, 123, 144, 164, 179, 196, 210, 223,
+  7, 15, 26, 37, 49, 63, 78, 96, 112, 129, 146, 166, 182, 200, 215, 228,
+  9, 19, 28, 39, 54, 69, 86, 102, 117, 132, 151, 170, 187, 206, 220, 230,
+  13, 24, 35, 46, 60, 73, 91, 108, 122, 137, 154, 174, 189, 207, 224, 235,
+  17, 30, 40, 53, 66, 82, 98, 115, 126, 142, 161, 180, 197, 213, 227, 237,
+  22, 36, 48, 62, 76, 92, 105, 120, 133, 147, 167, 186, 203, 219, 232, 240,
+  27, 44, 56, 70, 84, 99, 113, 127, 140, 156, 175, 193, 209, 226, 236, 244,
+  33, 51, 68, 79, 94, 110, 125, 138, 149, 162, 184, 202, 217, 229, 241, 247,
+  42, 61, 77, 90, 106, 121, 134, 148, 160, 173, 191, 211, 225, 238, 245, 251,
+  50, 72, 87, 100, 118, 128, 145, 158, 168, 183, 204, 222, 233, 242, 249, 253,
+  57, 80, 97, 111, 131, 143, 155, 169, 178, 192, 214, 231, 239, 246, 250, 254,
+  65, 88, 107, 124, 139, 152, 163, 177, 185, 199, 221, 234, 243, 248, 252, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_row_iscan_16x16[256]) = {
+  0, 1, 2, 4, 6, 9, 12, 17, 22, 29, 36, 43, 54, 64, 76, 86,
+  3, 5, 7, 11, 15, 19, 25, 32, 38, 48, 59, 68, 84, 99, 115, 130,
+  8, 10, 13, 18, 23, 27, 33, 42, 51, 60, 72, 88, 103, 119, 142, 167,
+  14, 16, 20, 26, 31, 37, 44, 53, 61, 73, 85, 100, 116, 135, 161, 185,
+  21, 24, 30, 35, 40, 47, 55, 65, 74, 81, 94, 112, 133, 154, 179, 205,
+  28, 34, 39, 45, 50, 58, 67, 77, 87, 96, 106, 121, 146, 169, 196, 212,
+  41, 46, 49, 56, 63, 70, 79, 90, 98, 107, 122, 138, 159, 182, 207, 222,
+  52, 57, 62, 69, 75, 83, 93, 102, 110, 120, 134, 150, 176, 195, 215, 226,
+  66, 71, 78, 82, 91, 97, 108, 113, 127, 136, 148, 168, 188, 202, 221, 232,
+  80, 89, 92, 101, 105, 114, 125, 131, 139, 151, 162, 177, 192, 208, 223, 234,
+  95, 104, 109, 117, 123, 128, 143, 144, 155, 165, 175, 190, 206, 219, 233, 239,
+  111, 118, 124, 129, 140, 147, 157, 164, 170, 181, 191, 203, 224, 230, 240,
+  243, 126, 132, 137, 145, 153, 160, 174, 178, 184, 197, 204, 216, 231, 237,
+  244, 246, 141, 149, 156, 166, 172, 180, 189, 199, 200, 210, 220, 228, 238,
+  242, 249, 251, 152, 163, 171, 183, 186, 193, 201, 211, 214, 218, 227, 236,
+  245, 247, 252, 253, 158, 173, 187, 194, 198, 209, 213, 217, 225, 229, 235,
+  241, 248, 250, 254, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_16x16[256]) = {
+  0, 2, 5, 9, 17, 24, 36, 44, 55, 72, 88, 104, 128, 143, 166, 179,
+  1, 4, 8, 13, 20, 30, 40, 54, 66, 79, 96, 113, 141, 154, 178, 196,
+  3, 7, 11, 18, 25, 33, 46, 57, 71, 86, 101, 119, 148, 164, 186, 201,
+  6, 12, 16, 23, 31, 39, 53, 64, 78, 92, 110, 127, 153, 169, 193, 208,
+  10, 14, 19, 28, 37, 47, 58, 67, 84, 98, 114, 133, 161, 176, 198, 214,
+  15, 21, 26, 34, 43, 52, 65, 77, 91, 106, 120, 140, 165, 185, 205, 221,
+  22, 27, 32, 41, 48, 60, 73, 85, 99, 116, 130, 151, 175, 190, 211, 225,
+  29, 35, 42, 49, 59, 69, 81, 95, 108, 125, 139, 155, 182, 197, 217, 229,
+  38, 45, 51, 61, 68, 80, 93, 105, 118, 134, 150, 168, 191, 207, 223, 234,
+  50, 56, 63, 74, 83, 94, 109, 117, 129, 147, 163, 177, 199, 213, 228, 238,
+  62, 70, 76, 87, 97, 107, 122, 131, 145, 159, 172, 188, 210, 222, 235, 242,
+  75, 82, 90, 102, 112, 124, 138, 146, 157, 173, 187, 202, 219, 230, 240, 245,
+  89, 100, 111, 123, 132, 142, 156, 167, 180, 189, 203, 216, 231, 237, 246, 250,
+  103, 115, 126, 136, 149, 162, 171, 183, 194, 204, 215, 224, 236, 241, 248,
+  252, 121, 135, 144, 158, 170, 181, 192, 200, 209, 218, 227, 233, 243, 244,
+  251, 254, 137, 152, 160, 174, 184, 195, 206, 212, 220, 226, 232, 239, 247,
+  249, 253, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, vp9_default_iscan_32x32[1024]) = {
+  0, 2, 5, 10, 17, 25, 38, 47, 62, 83, 101, 121, 145, 170, 193, 204,
+  210, 219, 229, 233, 245, 257, 275, 299, 342, 356, 377, 405, 455, 471, 495,
+  527, 1, 4, 8, 15, 22, 30, 45, 58, 74, 92, 112, 133, 158, 184, 203, 215, 222,
+  228, 234, 237, 256, 274, 298, 317, 355, 376, 404, 426, 470, 494, 526, 551,
+  3, 7, 12, 18, 28, 36, 52, 64, 82, 102, 118, 142, 164, 189, 208, 217, 224,
+  231, 235, 238, 273, 297, 316, 329, 375, 403, 425, 440, 493, 525, 550, 567,
+  6, 11, 16, 23, 31, 43, 60, 73, 90, 109, 126, 150, 173, 196, 211, 220, 226,
+  232, 236, 239, 296, 315, 328, 335, 402, 424, 439, 447, 524, 549, 566, 575,
+  9, 14, 19, 29, 37, 50, 65, 78, 95, 116, 134, 157, 179, 201, 214, 223, 244,
+  255, 272, 295, 341, 354, 374, 401, 454, 469, 492, 523, 582, 596, 617, 645,
+  13, 20, 26, 35, 44, 54, 72, 85, 105, 123, 140, 163, 182, 205, 216, 225,
+  254, 271, 294, 314, 353, 373, 400, 423, 468, 491, 522, 548, 595, 616, 644,
+  666, 21, 27, 33, 42, 53, 63, 80, 94, 113, 132, 151, 172, 190, 209, 218, 227,
+  270, 293, 313, 327, 372, 399, 422, 438, 490, 521, 547, 565, 615, 643, 665,
+  680, 24, 32, 39, 48, 57, 71, 88, 104, 120, 139, 159, 178, 197, 212, 221, 230,
+  292, 312, 326, 334, 398, 421, 437, 446, 520, 546, 564, 574, 642, 664, 679,
+  687, 34, 40, 46, 56, 68, 81, 96, 111, 130, 147, 167, 186, 243, 253, 269, 291,
+  340, 352, 371, 397, 453, 467, 489, 519, 581, 594, 614, 641, 693, 705, 723,
+  747, 41, 49, 55, 67, 77, 91, 107, 124, 138, 161, 177, 194, 252, 268, 290,
+  311, 351, 370, 396, 420, 466, 488, 518, 545, 593, 613, 640, 663, 704, 722,
+  746, 765, 51, 59, 66, 76, 89, 99, 119, 131, 149, 168, 181, 200, 267, 289,
+  310, 325, 369, 395, 419, 436, 487, 517, 544, 563, 612, 639, 662, 678, 721,
+  745, 764, 777, 61, 69, 75, 87, 100, 114, 129, 144, 162, 180, 191, 207, 288,
+  309, 324, 333, 394, 418, 435, 445, 516, 543, 562, 573, 638, 661, 677, 686,
+  744, 763, 776, 783, 70, 79, 86, 97, 108, 122, 137, 155, 242, 251, 266, 287,
+  339, 350, 368, 393, 452, 465, 486, 515, 580, 592, 611, 637, 692, 703, 720,
+  743, 788, 798, 813, 833, 84, 93, 103, 110, 125, 141, 154, 171, 250, 265, 286,
+  308, 349, 367, 392, 417, 464, 485, 514, 542, 591, 610, 636, 660, 702, 719,
+  742, 762, 797, 812, 832, 848, 98, 106, 115, 127, 143, 156, 169, 185, 264,
+  285, 307, 323, 366, 391, 416, 434, 484, 513, 541, 561, 609, 635, 659, 676,
+  718, 741, 761, 775, 811, 831, 847, 858, 117, 128, 136, 148, 160, 175, 188,
+  198, 284, 306, 322, 332, 390, 415, 433, 444, 512, 540, 560, 572, 634, 658,
+  675, 685, 740, 760, 774, 782, 830, 846, 857, 863, 135, 146, 152, 165, 241,
+  249, 263, 283, 338, 348, 365, 389, 451, 463, 483, 511, 579, 590, 608, 633,
+  691, 701, 717, 739, 787, 796, 810, 829, 867, 875, 887, 903, 153, 166, 174,
+  183, 248, 262, 282, 305, 347, 364, 388, 414, 462, 482, 510, 539, 589, 607,
+  632, 657, 700, 716, 738, 759, 795, 809, 828, 845, 874, 886, 902, 915, 176,
+  187, 195, 202, 261, 281, 304, 321, 363, 387, 413, 432, 481, 509, 538, 559,
+  606, 631, 656, 674, 715, 737, 758, 773, 808, 827, 844, 856, 885, 901, 914,
+  923, 192, 199, 206, 213, 280, 303, 320, 331, 386, 412, 431, 443, 508, 537,
+  558, 571, 630, 655, 673, 684, 736, 757, 772, 781, 826, 843, 855, 862, 900,
+  913, 922, 927, 240, 247, 260, 279, 337, 346, 362, 385, 450, 461, 480, 507,
+  578, 588, 605, 629, 690, 699, 714, 735, 786, 794, 807, 825, 866, 873, 884,
+  899, 930, 936, 945, 957, 246, 259, 278, 302, 345, 361, 384, 411, 460, 479,
+  506, 536, 587, 604, 628, 654, 698, 713, 734, 756, 793, 806, 824, 842, 872,
+  883, 898, 912, 935, 944, 956, 966, 258, 277, 301, 319, 360, 383, 410, 430,
+  478, 505, 535, 557, 603, 627, 653, 672, 712, 733, 755, 771, 805, 823, 841,
+  854, 882, 897, 911, 921, 943, 955, 965, 972, 276, 300, 318, 330, 382, 409,
+  429, 442, 504, 534, 556, 570, 626, 652, 671, 683, 732, 754, 770, 780, 822,
+  840, 853, 861, 896, 910, 920, 926, 954, 964, 971, 975, 336, 344, 359, 381,
+  449, 459, 477, 503, 577, 586, 602, 625, 689, 697, 711, 731, 785, 792, 804,
+  821, 865, 871, 881, 895, 929, 934, 942, 953, 977, 981, 987, 995, 343, 358,
+  380, 408, 458, 476, 502, 533, 585, 601, 624, 651, 696, 710, 730, 753, 791,
+  803, 820, 839, 870, 880, 894, 909, 933, 941, 952, 963, 980, 986, 994, 1001,
+  357, 379, 407, 428, 475, 501, 532, 555, 600, 623, 650, 670, 709, 729, 752,
+  769, 802, 819, 838, 852, 879, 893, 908, 919, 940, 951, 962, 970, 985, 993,
+  1000, 1005, 378, 406, 427, 441, 500, 531, 554, 569, 622, 649, 669, 682, 728,
+  751, 768, 779, 818, 837, 851, 860, 892, 907, 918, 925, 950, 961, 969, 974,
+  992, 999, 1004, 1007, 448, 457, 474, 499, 576, 584, 599, 621, 688, 695, 708,
+  727, 784, 790, 801, 817, 864, 869, 878, 891, 928, 932, 939, 949, 976, 979,
+  984, 991, 1008, 1010, 1013, 1017, 456, 473, 498, 530, 583, 598, 620, 648,
+  694, 707, 726, 750, 789, 800, 816, 836, 868, 877, 890, 906, 931, 938, 948,
+  960, 978, 983, 990, 998, 1009, 1012, 1016, 1020, 472, 497, 529, 553, 597,
+  619, 647, 668, 706, 725, 749, 767, 799, 815, 835, 850, 876, 889, 905, 917,
+  937, 947, 959, 968, 982, 989, 997, 1003, 1011, 1015, 1019, 1022, 496, 528,
+  552, 568, 618, 646, 667, 681, 724, 748, 766, 778, 814, 834, 849, 859, 888,
+  904, 916, 924, 946, 958, 967, 973, 988, 996, 1002, 1006, 1014, 1018, 1021,
+  1023,
+};
+
+const scan_order vp9_default_scan_orders[TX_SIZES] = {
+  {default_scan_4x4,   vp9_default_iscan_4x4,   default_scan_4x4_neighbors},
+  {default_scan_8x8,   vp9_default_iscan_8x8,   default_scan_8x8_neighbors},
+  {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors},
+  {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+};
+
+const scan_order vp9_scan_orders[TX_SIZES][TX_TYPES] = {
+  {  // TX_4X4
+    {default_scan_4x4, vp9_default_iscan_4x4, default_scan_4x4_neighbors},
+    {row_scan_4x4,     vp9_row_iscan_4x4,     row_scan_4x4_neighbors},
+    {col_scan_4x4,     vp9_col_iscan_4x4,     col_scan_4x4_neighbors},
+    {default_scan_4x4, vp9_default_iscan_4x4, default_scan_4x4_neighbors}
+  }, {  // TX_8X8
+    {default_scan_8x8, vp9_default_iscan_8x8, default_scan_8x8_neighbors},
+    {row_scan_8x8,     vp9_row_iscan_8x8,     row_scan_8x8_neighbors},
+    {col_scan_8x8,     vp9_col_iscan_8x8,     col_scan_8x8_neighbors},
+    {default_scan_8x8, vp9_default_iscan_8x8, default_scan_8x8_neighbors}
+  }, {  // TX_16X16
+    {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors},
+    {row_scan_16x16,     vp9_row_iscan_16x16,     row_scan_16x16_neighbors},
+    {col_scan_16x16,     vp9_col_iscan_16x16,     col_scan_16x16_neighbors},
+    {default_scan_16x16, vp9_default_iscan_16x16, default_scan_16x16_neighbors}
+  }, {  // TX_32X32
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+    {default_scan_32x32, vp9_default_iscan_32x32, default_scan_32x32_neighbors},
+  }
+};
diff --git a/media/libvpx/vp9/common/vp9_scan.h b/media/libvpx/vp9/common/vp9_scan.h
new file mode 100644
index 000000000..1d86b5cfe
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_scan.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SCAN_H_
+#define VP9_COMMON_VP9_SCAN_H_
+
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_enums.h"
+#include "vp9/common/vp9_blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_NEIGHBORS 2
+
+typedef struct {
+  const int16_t *scan;
+  const int16_t *iscan;
+  const int16_t *neighbors;
+} scan_order;
+
+extern const scan_order vp9_default_scan_orders[TX_SIZES];
+extern const scan_order vp9_scan_orders[TX_SIZES][TX_TYPES];
+
+static INLINE int get_coef_context(const int16_t *neighbors,
+                                   const uint8_t *token_cache, int c) {
+  return (1 + token_cache[neighbors[MAX_NEIGHBORS * c + 0]] +
+          token_cache[neighbors[MAX_NEIGHBORS * c + 1]]) >> 1;
+}
+
+static INLINE const scan_order *get_scan(const MACROBLOCKD *xd, TX_SIZE tx_size,
+                                         PLANE_TYPE type, int block_idx) {
+  const MODE_INFO *const mi = xd->mi[0];
+
+  if (is_inter_block(&mi->mbmi) || type != PLANE_TYPE_Y || xd->lossless) {
+    return &vp9_default_scan_orders[tx_size];
+  } else {
+    const PREDICTION_MODE mode = get_y_mode(mi, block_idx);
+    return &vp9_scan_orders[tx_size][intra_mode_to_tx_type_lookup[mode]];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SCAN_H_
diff --git a/media/libvpx/vp9/common/vp9_seg_common.c b/media/libvpx/vp9/common/vp9_seg_common.c
new file mode 100644
index 000000000..910200ecc
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_seg_common.c
@@ -0,0 +1,75 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_quant_common.h"
+
+static const int seg_feature_data_signed[SEG_LVL_MAX] = { 1, 1, 0, 0 };
+
+static const int seg_feature_data_max[SEG_LVL_MAX] = {
+  MAXQ, MAX_LOOP_FILTER, 3, 0 };
+
+// These functions provide access to new segment level features.
+// Eventually these function may be "optimized out" but for the moment,
+// the coding mechanism is still subject to change so these provide a
+// convenient single point of change.
+
+int vp9_segfeature_active(const struct segmentation *seg, int segment_id,
+                          SEG_LVL_FEATURES feature_id) {
+  return seg->enabled &&
+         (seg->feature_mask[segment_id] & (1 << feature_id));
+}
+
+void vp9_clearall_segfeatures(struct segmentation *seg) {
+  vp9_zero(seg->feature_data);
+  vp9_zero(seg->feature_mask);
+}
+
+void vp9_enable_segfeature(struct segmentation *seg, int segment_id,
+                           SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] |= 1 << feature_id;
+}
+
+int vp9_seg_feature_data_max(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_max[feature_id];
+}
+
+int vp9_is_segfeature_signed(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_signed[feature_id];
+}
+
+void vp9_set_segdata(struct segmentation *seg, int segment_id,
+                     SEG_LVL_FEATURES feature_id, int seg_data) {
+  assert(seg_data <= seg_feature_data_max[feature_id]);
+  if (seg_data < 0) {
+    assert(seg_feature_data_signed[feature_id]);
+    assert(-seg_data <= seg_feature_data_max[feature_id]);
+  }
+
+  seg->feature_data[segment_id][feature_id] = seg_data;
+}
+
+int vp9_get_segdata(const struct segmentation *seg, int segment_id,
+                    SEG_LVL_FEATURES feature_id) {
+  return seg->feature_data[segment_id][feature_id];
+}
+
+
+const vp9_tree_index vp9_segment_tree[TREE_SIZE(MAX_SEGMENTS)] = {
+  2,  4,  6,  8, 10, 12,
+  0, -1, -2, -3, -4, -5, -6, -7
+};
+
+
+// TBD? Functions to read and write segment data with range / validity checking
diff --git a/media/libvpx/vp9/common/vp9_seg_common.h b/media/libvpx/vp9/common/vp9_seg_common.h
new file mode 100644
index 000000000..ff2d66a36
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_seg_common.h
@@ -0,0 +1,82 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SEG_COMMON_H_
+#define VP9_COMMON_VP9_SEG_COMMON_H_
+
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define SEGMENT_DELTADATA   0
+#define SEGMENT_ABSDATA     1
+
+#define MAX_SEGMENTS     8
+#define SEG_TREE_PROBS   (MAX_SEGMENTS-1)
+
+#define PREDICTION_PROBS 3
+
+// Segment level features.
+typedef enum {
+  SEG_LVL_ALT_Q = 0,               // Use alternate Quantizer ....
+  SEG_LVL_ALT_LF = 1,              // Use alternate loop filter value...
+  SEG_LVL_REF_FRAME = 2,           // Optional Segment reference frame
+  SEG_LVL_SKIP = 3,                // Optional Segment (0,0) + skip mode
+  SEG_LVL_MAX = 4                  // Number of features supported
+} SEG_LVL_FEATURES;
+
+
+struct segmentation {
+  uint8_t enabled;
+  uint8_t update_map;
+  uint8_t update_data;
+  uint8_t abs_delta;
+  uint8_t temporal_update;
+
+  vp9_prob tree_probs[SEG_TREE_PROBS];
+  vp9_prob pred_probs[PREDICTION_PROBS];
+
+  int16_t feature_data[MAX_SEGMENTS][SEG_LVL_MAX];
+  unsigned int feature_mask[MAX_SEGMENTS];
+};
+
+int vp9_segfeature_active(const struct segmentation *seg,
+                          int segment_id,
+                          SEG_LVL_FEATURES feature_id);
+
+void vp9_clearall_segfeatures(struct segmentation *seg);
+
+void vp9_enable_segfeature(struct segmentation *seg,
+                           int segment_id,
+                           SEG_LVL_FEATURES feature_id);
+
+int vp9_seg_feature_data_max(SEG_LVL_FEATURES feature_id);
+
+int vp9_is_segfeature_signed(SEG_LVL_FEATURES feature_id);
+
+void vp9_set_segdata(struct segmentation *seg,
+                     int segment_id,
+                     SEG_LVL_FEATURES feature_id,
+                     int seg_data);
+
+int vp9_get_segdata(const struct segmentation *seg,
+                    int segment_id,
+                    SEG_LVL_FEATURES feature_id);
+
+extern const vp9_tree_index vp9_segment_tree[TREE_SIZE(MAX_SEGMENTS)];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SEG_COMMON_H_
+
diff --git a/media/libvpx/vp9/common/vp9_systemdependent.h b/media/libvpx/vp9/common/vp9_systemdependent.h
new file mode 100644
index 000000000..fc77762de
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_systemdependent.h
@@ -0,0 +1,84 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_SYSTEMDEPENDENT_H_
+#define VP9_COMMON_VP9_SYSTEMDEPENDENT_H_
+
+#include "vpx_ports/msvc.h"
+
+#ifdef _MSC_VER
+# include <math.h>  // the ceil() definition must precede intrin.h
+# if _MSC_VER > 1310 && (defined(_M_X64) || defined(_M_IX86))
+#  include <intrin.h>
+#  define USE_MSC_INTRINSICS
+# endif
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./vpx_config.h"
+#if ARCH_X86 || ARCH_X86_64
+void vpx_reset_mmx_state(void);
+#define vp9_clear_system_state() vpx_reset_mmx_state()
+#else
+#define vp9_clear_system_state()
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER < 1800
+// round is not defined in MSVC before VS2013.
+static INLINE int round(double x) {
+  if (x < 0)
+    return (int)ceil(x - 0.5);
+  else
+    return (int)floor(x + 0.5);
+}
+#endif
+
+// use GNU builtins where available.
+#if defined(__GNUC__) && \
+    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
+static INLINE int get_msb(unsigned int n) {
+  return 31 ^ __builtin_clz(n);
+}
+#elif defined(USE_MSC_INTRINSICS)
+#pragma intrinsic(_BitScanReverse)
+
+static INLINE int get_msb(unsigned int n) {
+  unsigned long first_set_bit;
+  _BitScanReverse(&first_set_bit, n);
+  return first_set_bit;
+}
+#undef USE_MSC_INTRINSICS
+#else
+// Returns (int)floor(log2(n)). n must be > 0.
+static INLINE int get_msb(unsigned int n) {
+  int log = 0;
+  unsigned int value = n;
+  int i;
+
+  for (i = 4; i >= 0; --i) {
+    const int shift = (1 << i);
+    const unsigned int x = value >> shift;
+    if (x != 0) {
+      value = x;
+      log += shift;
+    }
+  }
+  return log;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_SYSTEMDEPENDENT_H_
diff --git a/media/libvpx/vp9/common/vp9_thread.c b/media/libvpx/vp9/common/vp9_thread.c
new file mode 100644
index 000000000..1c6aec032
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_thread.c
@@ -0,0 +1,184 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Original source:
+//  http://git.chromium.org/webm/libwebp.git
+//  100644 blob 264210ba2807e4da47eb5d18c04cf869d89b9784  src/utils/thread.c
+
+#include <assert.h>
+#include <string.h>   // for memset()
+#include "./vp9_thread.h"
+#include "vpx_mem/vpx_mem.h"
+
+#if CONFIG_MULTITHREAD
+
+struct VP9WorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t  condition_;
+  pthread_t       thread_;
+};
+
+//------------------------------------------------------------------------------
+
+static void execute(VP9Worker *const worker);  // Forward declaration.
+
+static THREADFN thread_loop(void *ptr) {
+  VP9Worker *const worker = (VP9Worker*)ptr;
+  int done = 0;
+  while (!done) {
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    while (worker->status_ == OK) {   // wait in idling mode
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    if (worker->status_ == WORK) {
+      execute(worker);
+      worker->status_ = OK;
+    } else if (worker->status_ == NOT_OK) {   // finish the worker
+      done = 1;
+    }
+    // signal to the main thread that we're done (for sync())
+    pthread_cond_signal(&worker->impl_->condition_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+  }
+  return THREAD_RETURN(NULL);    // Thread is finished
+}
+
+// main thread state control
+static void change_state(VP9Worker *const worker,
+                         VP9WorkerStatus new_status) {
+  // No-op when attempting to change state on a thread that didn't come up.
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
+
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
+    // wait for the worker to finish
+    while (worker->status_ != OK) {
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    // assign new status and release the working thread if needed
+    if (new_status != OK) {
+      worker->status_ = new_status;
+      pthread_cond_signal(&worker->impl_->condition_);
+    }
+  }
+  pthread_mutex_unlock(&worker->impl_->mutex_);
+}
+
+#endif  // CONFIG_MULTITHREAD
+
+//------------------------------------------------------------------------------
+
+static void init(VP9Worker *const worker) {
+  memset(worker, 0, sizeof(*worker));
+  worker->status_ = NOT_OK;
+}
+
+static int sync(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, OK);
+#endif
+  assert(worker->status_ <= OK);
+  return !worker->had_error;
+}
+
+static int reset(VP9Worker *const worker) {
+  int ok = 1;
+  worker->had_error = 0;
+  if (worker->status_ < OK) {
+#if CONFIG_MULTITHREAD
+    worker->impl_ = (VP9WorkerImpl*)vpx_calloc(1, sizeof(*worker->impl_));
+    if (worker->impl_ == NULL) {
+      return 0;
+    }
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, thread_loop, worker);
+    if (ok) worker->status_ = OK;
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+ Error:
+      vpx_free(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
+#else
+    worker->status_ = OK;
+#endif
+  } else if (worker->status_ > OK) {
+    ok = sync(worker);
+  }
+  assert(!ok || (worker->status_ == OK));
+  return ok;
+}
+
+static void execute(VP9Worker *const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+static void launch(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, WORK);
+#else
+  execute(worker);
+#endif
+}
+
+static void end(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  if (worker->impl_ != NULL) {
+    change_state(worker, NOT_OK);
+    pthread_join(worker->impl_->thread_, NULL);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
+    pthread_cond_destroy(&worker->impl_->condition_);
+    vpx_free(worker->impl_);
+    worker->impl_ = NULL;
+  }
+#else
+  worker->status_ = NOT_OK;
+  assert(worker->impl_ == NULL);
+#endif
+  assert(worker->status_ == NOT_OK);
+}
+
+//------------------------------------------------------------------------------
+
+static VP9WorkerInterface g_worker_interface = {
+  init, reset, sync, launch, execute, end
+};
+
+int vp9_set_worker_interface(const VP9WorkerInterface* const winterface) {
+  if (winterface == NULL ||
+      winterface->init == NULL || winterface->reset == NULL ||
+      winterface->sync == NULL || winterface->launch == NULL ||
+      winterface->execute == NULL || winterface->end == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const VP9WorkerInterface *vp9_get_worker_interface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------
diff --git a/media/libvpx/vp9/common/vp9_thread.h b/media/libvpx/vp9/common/vp9_thread.h
new file mode 100644
index 000000000..12848fede
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_thread.h
@@ -0,0 +1,223 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// Multi-threaded worker
+//
+// Original source:
+//  http://git.chromium.org/webm/libwebp.git
+//  100644 blob 7bd451b124ae3b81596abfbcc823e3cb129d3a38  src/utils/thread.h
+
+#ifndef VP9_DECODER_VP9_THREAD_H_
+#define VP9_DECODER_VP9_THREAD_H_
+
+#include "./vpx_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Set maximum decode threads to be 8 due to the limit of frame buffers
+// and not enough semaphores in the emulation layer on windows.
+#define MAX_DECODE_THREADS 8
+
+#if CONFIG_MULTITHREAD
+
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+#include <errno.h>  // NOLINT
+#include <process.h>  // NOLINT
+#include <windows.h>  // NOLINT
+typedef HANDLE pthread_t;
+typedef CRITICAL_SECTION pthread_mutex_t;
+typedef struct {
+  HANDLE waiting_sem_;
+  HANDLE received_sem_;
+  HANDLE signal_event_;
+} pthread_cond_t;
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+// _beginthreadex requires __stdcall
+#define THREADFN unsigned int __stdcall
+#define THREAD_RETURN(val) (unsigned int)((DWORD_PTR)val)
+
+static INLINE int pthread_create(pthread_t* const thread, const void* attr,
+                                 unsigned int (__stdcall *start)(void*),
+                                 void* arg) {
+  (void)attr;
+  *thread = (pthread_t)_beginthreadex(NULL,   /* void *security */
+                                      0,      /* unsigned stack_size */
+                                      start,
+                                      arg,
+                                      0,      /* unsigned initflag */
+                                      NULL);  /* unsigned *thrdaddr */
+  if (*thread == NULL) return 1;
+  SetThreadPriority(*thread, THREAD_PRIORITY_ABOVE_NORMAL);
+  return 0;
+}
+
+static INLINE int pthread_join(pthread_t thread, void** value_ptr) {
+  (void)value_ptr;
+  return (WaitForSingleObject(thread, INFINITE) != WAIT_OBJECT_0 ||
+          CloseHandle(thread) == 0);
+}
+
+// Mutex
+static INLINE int pthread_mutex_init(pthread_mutex_t *const mutex,
+                                     void* mutexattr) {
+  (void)mutexattr;
+  InitializeCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_trylock(pthread_mutex_t *const mutex) {
+  return TryEnterCriticalSection(mutex) ? 0 : EBUSY;
+}
+
+static INLINE int pthread_mutex_lock(pthread_mutex_t *const mutex) {
+  EnterCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_unlock(pthread_mutex_t *const mutex) {
+  LeaveCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_destroy(pthread_mutex_t *const mutex) {
+  DeleteCriticalSection(mutex);
+  return 0;
+}
+
+// Condition
+static INLINE int pthread_cond_destroy(pthread_cond_t *const condition) {
+  int ok = 1;
+  ok &= (CloseHandle(condition->waiting_sem_) != 0);
+  ok &= (CloseHandle(condition->received_sem_) != 0);
+  ok &= (CloseHandle(condition->signal_event_) != 0);
+  return !ok;
+}
+
+static INLINE int pthread_cond_init(pthread_cond_t *const condition,
+                                    void* cond_attr) {
+  (void)cond_attr;
+  condition->waiting_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->received_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->signal_event_ = CreateEvent(NULL, FALSE, FALSE, NULL);
+  if (condition->waiting_sem_ == NULL ||
+      condition->received_sem_ == NULL ||
+      condition->signal_event_ == NULL) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+  return 0;
+}
+
+static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {
+  int ok = 1;
+  if (WaitForSingleObject(condition->waiting_sem_, 0) == WAIT_OBJECT_0) {
+    // a thread is waiting in pthread_cond_wait: allow it to be notified
+    ok = SetEvent(condition->signal_event_);
+    // wait until the event is consumed so the signaler cannot consume
+    // the event via its own pthread_cond_wait.
+    ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
+           WAIT_OBJECT_0);
+  }
+  return !ok;
+}
+
+static INLINE int pthread_cond_wait(pthread_cond_t *const condition,
+                                    pthread_mutex_t *const mutex) {
+  int ok;
+  // note that there is a consumer available so the signal isn't dropped in
+  // pthread_cond_signal
+  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL))
+    return 1;
+  // now unlock the mutex so pthread_cond_signal may be issued
+  pthread_mutex_unlock(mutex);
+  ok = (WaitForSingleObject(condition->signal_event_, INFINITE) ==
+        WAIT_OBJECT_0);
+  ok &= ReleaseSemaphore(condition->received_sem_, 1, NULL);
+  pthread_mutex_lock(mutex);
+  return !ok;
+}
+#else  // _WIN32
+#include <pthread.h> // NOLINT
+# define THREADFN void*
+# define THREAD_RETURN(val) val
+#endif
+
+#endif  // CONFIG_MULTITHREAD
+
+// State of the worker thread object
+typedef enum {
+  NOT_OK = 0,   // object is unusable
+  OK,           // ready to work
+  WORK          // busy finishing the current task
+} VP9WorkerStatus;
+
+// Function to be called by the worker thread. Takes two opaque pointers as
+// arguments (data1 and data2), and should return false in case of error.
+typedef int (*VP9WorkerHook)(void*, void*);
+
+// Platform-dependent implementation details for the worker.
+typedef struct VP9WorkerImpl VP9WorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
+typedef struct {
+  VP9WorkerImpl *impl_;
+  VP9WorkerStatus status_;
+  VP9WorkerHook hook;     // hook to call
+  void *data1;            // first argument passed to 'hook'
+  void *data2;            // second argument passed to 'hook'
+  int had_error;          // return value of the last call to 'hook'
+} VP9Worker;
+
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
+  // Must be called first, before any other method.
+  void (*init)(VP9Worker *const worker);
+  // Must be called to initialize the object and spawn the thread. Re-entrant.
+  // Will potentially launch the thread. Returns false in case of error.
+  int (*reset)(VP9Worker *const worker);
+  // Makes sure the previous work is finished. Returns true if worker->had_error
+  // was not set and no error condition was triggered by the working thread.
+  int (*sync)(VP9Worker *const worker);
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
+  // hook/data1/data2 values can be changed at any time before calling this
+  // function, but not be changed afterward until the next call to Sync().
+  void (*launch)(VP9Worker *const worker);
+  // This function is similar to launch() except that it calls the
+  // hook directly instead of using a thread. Convenient to bypass the thread
+  // mechanism while still using the VP9Worker structs. sync() must
+  // still be called afterward (for error reporting).
+  void (*execute)(VP9Worker *const worker);
+  // Kill the thread and terminate the object. To use the object again, one
+  // must call reset() again.
+  void (*end)(VP9Worker *const worker);
+} VP9WorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+int vp9_set_worker_interface(const VP9WorkerInterface *const winterface);
+
+// Retrieve the currently set thread worker interface.
+const VP9WorkerInterface *vp9_get_worker_interface(void);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_THREAD_H_
diff --git a/media/libvpx/vp9/common/vp9_thread_common.c b/media/libvpx/vp9/common/vp9_thread_common.c
new file mode 100644
index 000000000..cba57ff41
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_thread_common.c
@@ -0,0 +1,436 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_loopfilter.h"
+
+#if CONFIG_MULTITHREAD
+static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
+  const int kMaxTryLocks = 4000;
+  int locked = 0;
+  int i;
+
+  for (i = 0; i < kMaxTryLocks; ++i) {
+    if (!pthread_mutex_trylock(mutex)) {
+      locked = 1;
+      break;
+    }
+  }
+
+  if (!locked)
+    pthread_mutex_lock(mutex);
+}
+#endif  // CONFIG_MULTITHREAD
+
+static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+
+  if (r && !(c & (nsync - 1))) {
+    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+    mutex_lock(mutex);
+
+    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
+      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
+    }
+    pthread_mutex_unlock(mutex);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+#endif  // CONFIG_MULTITHREAD
+}
+
+static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
+                              const int sb_cols) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+  int cur;
+  // Only signal when there are enough filtered SB for next row to run.
+  int sig = 1;
+
+  if (c < sb_cols - 1) {
+    cur = c;
+    if (c % nsync)
+      sig = 0;
+  } else {
+    cur = sb_cols + nsync;
+  }
+
+  if (sig) {
+    mutex_lock(&lf_sync->mutex_[r]);
+
+    lf_sync->cur_sb_col[r] = cur;
+
+    pthread_cond_signal(&lf_sync->cond_[r]);
+    pthread_mutex_unlock(&lf_sync->mutex_[r]);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+  (void)sb_cols;
+#endif  // CONFIG_MULTITHREAD
+}
+
+// Implement row loopfiltering for each thread.
+static INLINE
+void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
+                             VP9_COMMON *const cm,
+                             struct macroblockd_plane planes[MAX_MB_PLANE],
+                             int start, int stop, int y_only,
+                             VP9LfSync *const lf_sync) {
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
+  int mi_row, mi_col;
+  enum lf_path path;
+  if (y_only)
+    path = LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    path = LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    path = LF_PATH_444;
+  else
+    path = LF_PATH_SLOW;
+
+  for (mi_row = start; mi_row < stop;
+       mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
+    MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
+      const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
+      const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
+      LOOP_FILTER_MASK lfm;
+      int plane;
+
+      sync_read(lf_sync, r, c);
+
+      vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
+
+      // TODO(JBB): Make setup_mask work for non 420.
+      vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
+                     &lfm);
+
+      vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
+      for (plane = 1; plane < num_planes; ++plane) {
+        switch (path) {
+          case LF_PATH_420:
+            vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_444:
+            vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_SLOW:
+            vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+            break;
+        }
+      }
+
+      sync_write(lf_sync, r, c, sb_cols);
+    }
+  }
+}
+
+// Row-based multi-threaded loopfilter hook
+static int loop_filter_row_worker(VP9LfSync *const lf_sync,
+                                  LFWorkerData *const lf_data) {
+  thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+                          lf_data->start, lf_data->stop, lf_data->y_only,
+                          lf_sync);
+  return 1;
+}
+
+static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
+                                VP9_COMMON *cm,
+                                struct macroblockd_plane planes[MAX_MB_PLANE],
+                                int start, int stop, int y_only,
+                                VP9Worker *workers, int nworkers,
+                                VP9LfSync *lf_sync) {
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+  // Number of superblock rows and cols
+  const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
+  // Decoder may allocate more threads than number of tiles based on user's
+  // input.
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int num_workers = MIN(nworkers, tile_cols);
+  int i;
+
+  if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
+      num_workers > lf_sync->num_workers) {
+    vp9_loop_filter_dealloc(lf_sync);
+    vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
+  }
+
+  // Initialize cur_sb_col to -1 for all SB rows.
+  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+  // Set up loopfilter thread data.
+  // The decoder is capping num_workers because it has been observed that using
+  // more threads on the loopfilter than there are cores will hurt performance
+  // on Android. This is because the system will only schedule the tile decode
+  // workers on cores equal to the number of tile columns. Then if the decoder
+  // tries to use more threads for the loopfilter, it will hurt performance
+  // because of contention. If the multithreading code changes in the future
+  // then the number of workers used by the loopfilter should be revisited.
+  for (i = 0; i < num_workers; ++i) {
+    VP9Worker *const worker = &workers[i];
+    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+    worker->hook = (VP9WorkerHook)loop_filter_row_worker;
+    worker->data1 = lf_sync;
+    worker->data2 = lf_data;
+
+    // Loopfilter data
+    vp9_loop_filter_data_reset(lf_data, frame, cm, planes);
+    lf_data->start = start + i * MI_BLOCK_SIZE;
+    lf_data->stop = stop;
+    lf_data->y_only = y_only;
+
+    // Start loopfiltering
+    if (i == num_workers - 1) {
+      winterface->execute(worker);
+    } else {
+      winterface->launch(worker);
+    }
+  }
+
+  // Wait till all rows are finished
+  for (i = 0; i < num_workers; ++i) {
+    winterface->sync(&workers[i]);
+  }
+}
+
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              VP9_COMMON *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level,
+                              int y_only, int partial_frame,
+                              VP9Worker *workers, int num_workers,
+                              VP9LfSync *lf_sync) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+
+  if (!frame_filter_level) return;
+
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = MAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  vp9_loop_filter_frame_init(cm, frame_filter_level);
+
+  loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
+                      y_only, workers, num_workers, lf_sync);
+}
+
+// Set up nsync by width.
+static INLINE int get_sync_range(int width) {
+  // nsync numbers are picked by testing. For example, for 4k
+  // video, using 4 gives best performance.
+  if (width < 640)
+    return 1;
+  else if (width <= 1280)
+    return 2;
+  else if (width <= 4096)
+    return 4;
+  else
+    return 8;
+}
+
+// Allocate memory for lf row synchronization
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
+                           int width, int num_workers) {
+  lf_sync->rows = rows;
+#if CONFIG_MULTITHREAD
+  {
+    int i;
+
+    CHECK_MEM_ERROR(cm, lf_sync->mutex_,
+                    vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
+    if (lf_sync->mutex_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_mutex_init(&lf_sync->mutex_[i], NULL);
+      }
+    }
+
+    CHECK_MEM_ERROR(cm, lf_sync->cond_,
+                    vpx_malloc(sizeof(*lf_sync->cond_) * rows));
+    if (lf_sync->cond_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_cond_init(&lf_sync->cond_[i], NULL);
+      }
+    }
+  }
+#endif  // CONFIG_MULTITHREAD
+
+  CHECK_MEM_ERROR(cm, lf_sync->lfdata,
+                  vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
+  lf_sync->num_workers = num_workers;
+
+  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
+                  vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
+
+  // Set up nsync.
+  lf_sync->sync_range = get_sync_range(width);
+}
+
+// Deallocate lf synchronization related mutex and data
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) {
+  if (lf_sync != NULL) {
+#if CONFIG_MULTITHREAD
+    int i;
+
+    if (lf_sync->mutex_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_mutex_destroy(&lf_sync->mutex_[i]);
+      }
+      vpx_free(lf_sync->mutex_);
+    }
+    if (lf_sync->cond_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_cond_destroy(&lf_sync->cond_[i]);
+      }
+      vpx_free(lf_sync->cond_);
+    }
+#endif  // CONFIG_MULTITHREAD
+    vpx_free(lf_sync->lfdata);
+    vpx_free(lf_sync->cur_sb_col);
+    // clear the structure as the source of this call may be a resize in which
+    // case this call will be followed by an _alloc() which may fail.
+    vp9_zero(*lf_sync);
+  }
+}
+
+// Accumulate frame counts.
+void vp9_accumulate_frame_counts(VP9_COMMON *cm, FRAME_COUNTS *counts,
+                                 int is_dec) {
+  int i, j, k, l, m;
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
+    for (j = 0; j < INTRA_MODES; j++)
+      cm->counts.y_mode[i][j] += counts->y_mode[i][j];
+
+  for (i = 0; i < INTRA_MODES; i++)
+    for (j = 0; j < INTRA_MODES; j++)
+      cm->counts.uv_mode[i][j] += counts->uv_mode[i][j];
+
+  for (i = 0; i < PARTITION_CONTEXTS; i++)
+    for (j = 0; j < PARTITION_TYPES; j++)
+      cm->counts.partition[i][j] += counts->partition[i][j];
+
+  if (is_dec) {
+    int n;
+    for (i = 0; i < TX_SIZES; i++)
+      for (j = 0; j < PLANE_TYPES; j++)
+        for (k = 0; k < REF_TYPES; k++)
+          for (l = 0; l < COEF_BANDS; l++)
+            for (m = 0; m < COEFF_CONTEXTS; m++) {
+              cm->counts.eob_branch[i][j][k][l][m] +=
+                  counts->eob_branch[i][j][k][l][m];
+              for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
+                cm->counts.coef[i][j][k][l][m][n] +=
+                    counts->coef[i][j][k][l][m][n];
+            }
+  } else {
+    for (i = 0; i < TX_SIZES; i++)
+      for (j = 0; j < PLANE_TYPES; j++)
+        for (k = 0; k < REF_TYPES; k++)
+          for (l = 0; l < COEF_BANDS; l++)
+            for (m = 0; m < COEFF_CONTEXTS; m++)
+              cm->counts.eob_branch[i][j][k][l][m] +=
+                  counts->eob_branch[i][j][k][l][m];
+                // In the encoder, cm->counts.coef is only updated at frame
+                // level, so not need to accumulate it here.
+                // for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
+                //   cm->counts.coef[i][j][k][l][m][n] +=
+                //       counts->coef[i][j][k][l][m][n];
+  }
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+    for (j = 0; j < SWITCHABLE_FILTERS; j++)
+      cm->counts.switchable_interp[i][j] += counts->switchable_interp[i][j];
+
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    for (j = 0; j < INTER_MODES; j++)
+      cm->counts.inter_mode[i][j] += counts->inter_mode[i][j];
+
+  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      cm->counts.intra_inter[i][j] += counts->intra_inter[i][j];
+
+  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      cm->counts.comp_inter[i][j] += counts->comp_inter[i][j];
+
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      for (k = 0; k < 2; k++)
+      cm->counts.single_ref[i][j][k] += counts->single_ref[i][j][k];
+
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      cm->counts.comp_ref[i][j] += counts->comp_ref[i][j];
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
+    for (j = 0; j < TX_SIZES; j++)
+      cm->counts.tx.p32x32[i][j] += counts->tx.p32x32[i][j];
+
+    for (j = 0; j < TX_SIZES - 1; j++)
+      cm->counts.tx.p16x16[i][j] += counts->tx.p16x16[i][j];
+
+    for (j = 0; j < TX_SIZES - 2; j++)
+      cm->counts.tx.p8x8[i][j] += counts->tx.p8x8[i][j];
+  }
+
+  for (i = 0; i < TX_SIZES; i++)
+    cm->counts.tx.tx_totals[i] += counts->tx.tx_totals[i];
+
+  for (i = 0; i < SKIP_CONTEXTS; i++)
+    for (j = 0; j < 2; j++)
+      cm->counts.skip[i][j] += counts->skip[i][j];
+
+  for (i = 0; i < MV_JOINTS; i++)
+    cm->counts.mv.joints[i] += counts->mv.joints[i];
+
+  for (k = 0; k < 2; k++) {
+    nmv_component_counts *comps = &cm->counts.mv.comps[k];
+    nmv_component_counts *comps_t = &counts->mv.comps[k];
+
+    for (i = 0; i < 2; i++) {
+      comps->sign[i] += comps_t->sign[i];
+      comps->class0_hp[i] += comps_t->class0_hp[i];
+      comps->hp[i] += comps_t->hp[i];
+    }
+
+    for (i = 0; i < MV_CLASSES; i++)
+      comps->classes[i] += comps_t->classes[i];
+
+    for (i = 0; i < CLASS0_SIZE; i++) {
+      comps->class0[i] += comps_t->class0[i];
+      for (j = 0; j < MV_FP_SIZE; j++)
+        comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
+    }
+
+    for (i = 0; i < MV_OFFSET_BITS; i++)
+      for (j = 0; j < 2; j++)
+        comps->bits[i][j] += comps_t->bits[i][j];
+
+    for (i = 0; i < MV_FP_SIZE; i++)
+      comps->fp[i] += comps_t->fp[i];
+  }
+}
diff --git a/media/libvpx/vp9/common/vp9_thread_common.h b/media/libvpx/vp9/common/vp9_thread_common.h
new file mode 100644
index 000000000..3b3a6996a
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_thread_common.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
+#define VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
+#include "./vpx_config.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_thread.h"
+
+struct VP9Common;
+struct FRAME_COUNTS;
+
+// Loopfilter row synchronization
+typedef struct VP9LfSyncData {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t *mutex_;
+  pthread_cond_t *cond_;
+#endif
+  // Allocate memory to store the loop-filtered superblock index in each row.
+  int *cur_sb_col;
+  // The optimal sync_range for different resolution and platform should be
+  // determined by testing. Currently, it is chosen to be a power-of-2 number.
+  int sync_range;
+  int rows;
+
+  // Row-based parallel loopfilter data
+  LFWorkerData *lfdata;
+  int num_workers;
+} VP9LfSync;
+
+// Allocate memory for loopfilter row synchronization.
+void vp9_loop_filter_alloc(VP9LfSync *lf_sync, struct VP9Common *cm, int rows,
+                           int width, int num_workers);
+
+// Deallocate loopfilter synchronization related mutex and data.
+void vp9_loop_filter_dealloc(VP9LfSync *lf_sync);
+
+// Multi-threaded loopfilter that uses the tile threads.
+void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+                              struct VP9Common *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level,
+                              int y_only, int partial_frame,
+                              VP9Worker *workers, int num_workers,
+                              VP9LfSync *lf_sync);
+
+void vp9_accumulate_frame_counts(struct VP9Common *cm,
+                                 struct FRAME_COUNTS *counts, int is_dec);
+
+#endif  // VP9_COMMON_VP9_LOOPFILTER_THREAD_H_
diff --git a/media/libvpx/vp9/common/vp9_tile_common.c b/media/libvpx/vp9/common/vp9_tile_common.c
new file mode 100644
index 000000000..7a20e0a9e
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_tile_common.c
@@ -0,0 +1,59 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/common/vp9_onyxc_int.h"
+
+#define MIN_TILE_WIDTH_B64 4
+#define MAX_TILE_WIDTH_B64 64
+
+static int get_tile_offset(int idx, int mis, int log2) {
+  const int sb_cols = mi_cols_aligned_to_sb(mis) >> MI_BLOCK_SIZE_LOG2;
+  const int offset = ((idx * sb_cols) >> log2) << MI_BLOCK_SIZE_LOG2;
+  return MIN(offset, mis);
+}
+
+void vp9_tile_set_row(TileInfo *tile, const VP9_COMMON *cm, int row) {
+  tile->mi_row_start = get_tile_offset(row, cm->mi_rows, cm->log2_tile_rows);
+  tile->mi_row_end = get_tile_offset(row + 1, cm->mi_rows, cm->log2_tile_rows);
+}
+
+void vp9_tile_set_col(TileInfo *tile, const VP9_COMMON *cm, int col) {
+  tile->mi_col_start = get_tile_offset(col, cm->mi_cols, cm->log2_tile_cols);
+  tile->mi_col_end = get_tile_offset(col + 1, cm->mi_cols, cm->log2_tile_cols);
+}
+
+void vp9_tile_init(TileInfo *tile, const VP9_COMMON *cm, int row, int col) {
+  vp9_tile_set_row(tile, cm, row);
+  vp9_tile_set_col(tile, cm, col);
+}
+
+static int get_min_log2_tile_cols(const int sb64_cols) {
+  int min_log2 = 0;
+  while ((MAX_TILE_WIDTH_B64 << min_log2) < sb64_cols)
+    ++min_log2;
+  return min_log2;
+}
+
+static int get_max_log2_tile_cols(const int sb64_cols) {
+  int max_log2 = 1;
+  while ((sb64_cols >> max_log2) >= MIN_TILE_WIDTH_B64)
+    ++max_log2;
+  return max_log2 - 1;
+}
+
+void vp9_get_tile_n_bits(int mi_cols,
+                         int *min_log2_tile_cols, int *max_log2_tile_cols) {
+  const int sb64_cols = mi_cols_aligned_to_sb(mi_cols) >> MI_BLOCK_SIZE_LOG2;
+  *min_log2_tile_cols = get_min_log2_tile_cols(sb64_cols);
+  *max_log2_tile_cols = get_max_log2_tile_cols(sb64_cols);
+  assert(*min_log2_tile_cols <= *max_log2_tile_cols);
+}
diff --git a/media/libvpx/vp9/common/vp9_tile_common.h b/media/libvpx/vp9/common/vp9_tile_common.h
new file mode 100644
index 000000000..ae58805de
--- /dev/null
+++ b/media/libvpx/vp9/common/vp9_tile_common.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_COMMON_VP9_TILE_COMMON_H_
+#define VP9_COMMON_VP9_TILE_COMMON_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+
+typedef struct TileInfo {
+  int mi_row_start, mi_row_end;
+  int mi_col_start, mi_col_end;
+} TileInfo;
+
+// initializes 'tile->mi_(row|col)_(start|end)' for (row, col) based on
+// 'cm->log2_tile_(rows|cols)' & 'cm->mi_(rows|cols)'
+void vp9_tile_init(TileInfo *tile, const struct VP9Common *cm,
+                   int row, int col);
+
+void vp9_tile_set_row(TileInfo *tile, const struct VP9Common *cm, int row);
+void vp9_tile_set_col(TileInfo *tile, const struct VP9Common *cm, int col);
+
+void vp9_get_tile_n_bits(int mi_cols,
+                         int *min_log2_tile_cols, int *max_log2_tile_cols);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_COMMON_VP9_TILE_COMMON_H_
diff --git a/media/libvpx/vp9/common/x86/convolve.h b/media/libvpx/vp9/common/x86/convolve.h
new file mode 100644
index 000000000..de2df47e5
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/convolve.h
@@ -0,0 +1,296 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#ifndef VP9_COMMON_X86_CONVOLVE_H_
+#define VP9_COMMON_X86_CONVOLVE_H_
+
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_ports/mem.h"
+
+typedef void filter8_1dfunction (
+  const uint8_t *src_ptr,
+  ptrdiff_t src_pitch,
+  uint8_t *output_ptr,
+  ptrdiff_t out_pitch,
+  uint32_t output_height,
+  const int16_t *filter
+);
+
+#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \
+  void vp9_convolve8_##name##_##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                                    uint8_t *dst, ptrdiff_t dst_stride, \
+                                    const int16_t *filter_x, int x_step_q4, \
+                                    const int16_t *filter_y, int y_step_q4, \
+                                    int w, int h) { \
+  if (step_q4 == 16 && filter[3] != 128) { \
+    if (filter[0] || filter[1] || filter[2]) { \
+      while (w >= 16) { \
+        vp9_filter_block1d16_##dir##8_##avg##opt(src_start, \
+                                                 src_stride, \
+                                                 dst, \
+                                                 dst_stride, \
+                                                 h, \
+                                                 filter); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vp9_filter_block1d8_##dir##8_##avg##opt(src_start, \
+                                                src_stride, \
+                                                dst, \
+                                                dst_stride, \
+                                                h, \
+                                                filter); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vp9_filter_block1d4_##dir##8_##avg##opt(src_start, \
+                                                src_stride, \
+                                                dst, \
+                                                dst_stride, \
+                                                h, \
+                                                filter); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } else { \
+      while (w >= 16) { \
+        vp9_filter_block1d16_##dir##2_##avg##opt(src, \
+                                                 src_stride, \
+                                                 dst, \
+                                                 dst_stride, \
+                                                 h, \
+                                                 filter); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vp9_filter_block1d8_##dir##2_##avg##opt(src, \
+                                                src_stride, \
+                                                dst, \
+                                                dst_stride, \
+                                                h, \
+                                                filter); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vp9_filter_block1d4_##dir##2_##avg##opt(src, \
+                                                src_stride, \
+                                                dst, \
+                                                dst_stride, \
+                                                h, \
+                                                filter); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } \
+  } \
+  if (w) { \
+    vp9_convolve8_##name##_c(src, src_stride, dst, dst_stride, \
+                             filter_x, x_step_q4, filter_y, y_step_q4, \
+                             w, h); \
+  } \
+}
+
+#define FUN_CONV_2D(avg, opt) \
+void vp9_convolve8_##avg##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                              uint8_t *dst, ptrdiff_t dst_stride, \
+                              const int16_t *filter_x, int x_step_q4, \
+                              const int16_t *filter_y, int y_step_q4, \
+                              int w, int h) { \
+  assert(w <= 64); \
+  assert(h <= 64); \
+  if (x_step_q4 == 16 && y_step_q4 == 16) { \
+    if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 || \
+        filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) { \
+      DECLARE_ALIGNED(16, uint8_t, fdata2[64 * 71]); \
+      vp9_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, fdata2, 64, \
+                                filter_x, x_step_q4, filter_y, y_step_q4, \
+                                w, h + 7); \
+      vp9_convolve8_##avg##vert_##opt(fdata2 + 3 * 64, 64, dst, dst_stride, \
+                                      filter_x, x_step_q4, filter_y, \
+                                      y_step_q4, w, h); \
+    } else { \
+      DECLARE_ALIGNED(16, uint8_t, fdata2[64 * 65]); \
+      vp9_convolve8_horiz_##opt(src, src_stride, fdata2, 64, \
+                                filter_x, x_step_q4, filter_y, y_step_q4, \
+                                w, h + 1); \
+      vp9_convolve8_##avg##vert_##opt(fdata2, 64, dst, dst_stride, \
+                                      filter_x, x_step_q4, filter_y, \
+                                      y_step_q4, w, h); \
+    } \
+  } else { \
+    vp9_convolve8_##avg##c(src, src_stride, dst, dst_stride, \
+                           filter_x, x_step_q4, filter_y, y_step_q4, w, h); \
+  } \
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+
+typedef void highbd_filter8_1dfunction (
+  const uint16_t *src_ptr,
+  const ptrdiff_t src_pitch,
+  uint16_t *output_ptr,
+  ptrdiff_t out_pitch,
+  unsigned int output_height,
+  const int16_t *filter,
+  int bd
+);
+
+#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \
+  void vp9_highbd_convolve8_##name##_##opt(const uint8_t *src8, \
+                                           ptrdiff_t src_stride, \
+                                           uint8_t *dst8, \
+                                           ptrdiff_t dst_stride, \
+                                           const int16_t *filter_x, \
+                                           int x_step_q4, \
+                                           const int16_t *filter_y, \
+                                           int y_step_q4, \
+                                           int w, int h, int bd) { \
+  if (step_q4 == 16 && filter[3] != 128) { \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+    if (filter[0] || filter[1] || filter[2]) { \
+      while (w >= 16) { \
+        vp9_highbd_filter_block1d16_##dir##8_##avg##opt(src_start, \
+                                                        src_stride, \
+                                                        dst, \
+                                                        dst_stride, \
+                                                        h, \
+                                                        filter, \
+                                                        bd); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vp9_highbd_filter_block1d8_##dir##8_##avg##opt(src_start, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vp9_highbd_filter_block1d4_##dir##8_##avg##opt(src_start, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } else { \
+      while (w >= 16) { \
+        vp9_highbd_filter_block1d16_##dir##2_##avg##opt(src, \
+                                                        src_stride, \
+                                                        dst, \
+                                                        dst_stride, \
+                                                        h, \
+                                                        filter, \
+                                                        bd); \
+        src += 16; \
+        dst += 16; \
+        w -= 16; \
+      } \
+      while (w >= 8) { \
+        vp9_highbd_filter_block1d8_##dir##2_##avg##opt(src, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 8; \
+        dst += 8; \
+        w -= 8; \
+      } \
+      while (w >= 4) { \
+        vp9_highbd_filter_block1d4_##dir##2_##avg##opt(src, \
+                                                       src_stride, \
+                                                       dst, \
+                                                       dst_stride, \
+                                                       h, \
+                                                       filter, \
+                                                       bd); \
+        src += 4; \
+        dst += 4; \
+        w -= 4; \
+      } \
+    } \
+  } \
+  if (w) { \
+    vp9_highbd_convolve8_##name##_c(src8, src_stride, dst8, dst_stride, \
+                                    filter_x, x_step_q4, filter_y, y_step_q4, \
+                                    w, h, bd); \
+  } \
+}
+
+#define HIGH_FUN_CONV_2D(avg, opt) \
+void vp9_highbd_convolve8_##avg##opt(const uint8_t *src, ptrdiff_t src_stride, \
+                                     uint8_t *dst, ptrdiff_t dst_stride, \
+                                     const int16_t *filter_x, int x_step_q4, \
+                                     const int16_t *filter_y, int y_step_q4, \
+                                     int w, int h, int bd) { \
+  assert(w <= 64); \
+  assert(h <= 64); \
+  if (x_step_q4 == 16 && y_step_q4 == 16) { \
+    if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 || \
+        filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) { \
+      DECLARE_ALIGNED(16, uint16_t, fdata2[64 * 71]); \
+      vp9_highbd_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, \
+                                       CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                       filter_x, x_step_q4, \
+                                       filter_y, y_step_q4, \
+                                       w, h + 7, bd); \
+      vp9_highbd_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2) + 192, \
+                                             64, dst, dst_stride, \
+                                             filter_x, x_step_q4, \
+                                             filter_y, y_step_q4, \
+                                             w, h, bd); \
+    } else { \
+      DECLARE_ALIGNED(16, uint16_t, fdata2[64 * 65]); \
+      vp9_highbd_convolve8_horiz_##opt(src, src_stride, \
+                                       CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                       filter_x, x_step_q4, \
+                                       filter_y, y_step_q4, \
+                                       w, h + 1, bd); \
+      vp9_highbd_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2), 64, \
+                                             dst, dst_stride, \
+                                             filter_x, x_step_q4, \
+                                             filter_y, y_step_q4, \
+                                             w, h, bd); \
+    } \
+  } else { \
+    vp9_highbd_convolve8_##avg##c(src, src_stride, dst, dst_stride, \
+                                  filter_x, x_step_q4, filter_y, y_step_q4, w, \
+                                  h, bd); \
+  } \
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#endif  // VP9_COMMON_X86_CONVOLVE_H_
diff --git a/media/libvpx/vp9/common/x86/vp9_asm_stubs.c b/media/libvpx/vp9/common/x86/vp9_asm_stubs.c
new file mode 100644
index 000000000..fd55fb8c6
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_asm_stubs.c
@@ -0,0 +1,162 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/x86/convolve.h"
+
+#if HAVE_SSE2
+filter8_1dfunction vp9_filter_block1d16_v8_sse2;
+filter8_1dfunction vp9_filter_block1d16_h8_sse2;
+filter8_1dfunction vp9_filter_block1d8_v8_sse2;
+filter8_1dfunction vp9_filter_block1d8_h8_sse2;
+filter8_1dfunction vp9_filter_block1d4_v8_sse2;
+filter8_1dfunction vp9_filter_block1d4_h8_sse2;
+filter8_1dfunction vp9_filter_block1d16_v8_avg_sse2;
+filter8_1dfunction vp9_filter_block1d16_h8_avg_sse2;
+filter8_1dfunction vp9_filter_block1d8_v8_avg_sse2;
+filter8_1dfunction vp9_filter_block1d8_h8_avg_sse2;
+filter8_1dfunction vp9_filter_block1d4_v8_avg_sse2;
+filter8_1dfunction vp9_filter_block1d4_h8_avg_sse2;
+
+filter8_1dfunction vp9_filter_block1d16_v2_sse2;
+filter8_1dfunction vp9_filter_block1d16_h2_sse2;
+filter8_1dfunction vp9_filter_block1d8_v2_sse2;
+filter8_1dfunction vp9_filter_block1d8_h2_sse2;
+filter8_1dfunction vp9_filter_block1d4_v2_sse2;
+filter8_1dfunction vp9_filter_block1d4_h2_sse2;
+filter8_1dfunction vp9_filter_block1d16_v2_avg_sse2;
+filter8_1dfunction vp9_filter_block1d16_h2_avg_sse2;
+filter8_1dfunction vp9_filter_block1d8_v2_avg_sse2;
+filter8_1dfunction vp9_filter_block1d8_h2_avg_sse2;
+filter8_1dfunction vp9_filter_block1d4_v2_avg_sse2;
+filter8_1dfunction vp9_filter_block1d4_h2_avg_sse2;
+
+// void vp9_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void vp9_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+// void vp9_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+// void vp9_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                  uint8_t *dst, ptrdiff_t dst_stride,
+//                                  const int16_t *filter_x, int x_step_q4,
+//                                  const int16_t *filter_y, int y_step_q4,
+//                                  int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, sse2);
+
+// void vp9_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                         uint8_t *dst, ptrdiff_t dst_stride,
+//                         const int16_t *filter_x, int x_step_q4,
+//                         const int16_t *filter_y, int y_step_q4,
+//                         int w, int h);
+// void vp9_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                             uint8_t *dst, ptrdiff_t dst_stride,
+//                             const int16_t *filter_x, int x_step_q4,
+//                             const int16_t *filter_y, int y_step_q4,
+//                             int w, int h);
+FUN_CONV_2D(, sse2);
+FUN_CONV_2D(avg_ , sse2);
+
+#if CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_v8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_h8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_v8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_h8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_v8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_h8_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_v8_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_h8_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_v8_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_h8_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_v8_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_h8_avg_sse2;
+
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_v2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_h2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_v2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_h2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_v2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_h2_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_v2_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d16_h2_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_v2_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d8_h2_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_v2_avg_sse2;
+highbd_filter8_1dfunction vp9_highbd_filter_block1d4_h2_avg_sse2;
+
+// void vp9_highbd_convolve8_horiz_sse2(const uint8_t *src,
+//                                      ptrdiff_t src_stride,
+//                                      uint8_t *dst,
+//                                      ptrdiff_t dst_stride,
+//                                      const int16_t *filter_x,
+//                                      int x_step_q4,
+//                                      const int16_t *filter_y,
+//                                      int y_step_q4,
+//                                      int w, int h, int bd);
+// void vp9_highbd_convolve8_vert_sse2(const uint8_t *src,
+//                                     ptrdiff_t src_stride,
+//                                     uint8_t *dst,
+//                                     ptrdiff_t dst_stride,
+//                                     const int16_t *filter_x,
+//                                     int x_step_q4,
+//                                     const int16_t *filter_y,
+//                                     int y_step_q4,
+//                                     int w, int h, int bd);
+// void vp9_highbd_convolve8_avg_horiz_sse2(const uint8_t *src,
+//                                          ptrdiff_t src_stride,
+//                                          uint8_t *dst,
+//                                          ptrdiff_t dst_stride,
+//                                          const int16_t *filter_x,
+//                                          int x_step_q4,
+//                                          const int16_t *filter_y,
+//                                          int y_step_q4,
+//                                          int w, int h, int bd);
+// void vp9_highbd_convolve8_avg_vert_sse2(const uint8_t *src,
+//                                         ptrdiff_t src_stride,
+//                                         uint8_t *dst,
+//                                         ptrdiff_t dst_stride,
+//                                         const int16_t *filter_x,
+//                                         int x_step_q4,
+//                                         const int16_t *filter_y,
+//                                         int y_step_q4,
+//                                         int w, int h, int bd);
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+                 sse2);
+
+// void vp9_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h, int bd);
+// void vp9_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h, int bd);
+HIGH_FUN_CONV_2D(, sse2);
+HIGH_FUN_CONV_2D(avg_ , sse2);
+#endif  // CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
+#endif  // HAVE_SSE2
diff --git a/media/libvpx/vp9/common/x86/vp9_copy_sse2.asm b/media/libvpx/vp9/common/x86/vp9_copy_sse2.asm
new file mode 100644
index 000000000..b26383708
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_copy_sse2.asm
@@ -0,0 +1,156 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro convolve_fn 1
+INIT_XMM sse2
+cglobal convolve_%1, 4, 7, 4, src, src_stride, dst, dst_stride, \
+                              fx, fxs, fy, fys, w, h
+  mov r4d, dword wm
+  cmp r4d, 4
+  je .w4
+  cmp r4d, 8
+  je .w8
+  cmp r4d, 16
+  je .w16
+  cmp r4d, 32
+  je .w32
+
+  mov                    r4d, dword hm
+.loop64:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavgb                   m0, [dstq]
+  pavgb                   m1, [dstq+16]
+  pavgb                   m2, [dstq+32]
+  pavgb                   m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  add                   dstq, dst_strideq
+  dec                    r4d
+  jnz .loop64
+  RET
+
+.w32:
+  mov                    r4d, dword hm
+.loop32:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+src_strideq]
+  movu                    m3, [srcq+src_strideq+16]
+  lea                   srcq, [srcq+src_strideq*2]
+%ifidn %1, avg
+  pavgb                   m0, [dstq]
+  pavgb                   m1, [dstq            +16]
+  pavgb                   m2, [dstq+dst_strideq]
+  pavgb                   m3, [dstq+dst_strideq+16]
+%endif
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq+dst_strideq   ], m2
+  mova [dstq+dst_strideq+16], m3
+  lea                   dstq, [dstq+dst_strideq*2]
+  sub                    r4d, 2
+  jnz .loop32
+  RET
+
+.w16:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop16:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+src_strideq]
+  movu                    m2, [srcq+src_strideq*2]
+  movu                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  pavgb                   m0, [dstq]
+  pavgb                   m1, [dstq+dst_strideq]
+  pavgb                   m2, [dstq+dst_strideq*2]
+  pavgb                   m3, [dstq+r6q]
+%endif
+  mova  [dstq              ], m0
+  mova  [dstq+dst_strideq  ], m1
+  mova  [dstq+dst_strideq*2], m2
+  mova  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop16
+  RET
+
+INIT_MMX sse
+.w8:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop8:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+src_strideq]
+  movu                    m2, [srcq+src_strideq*2]
+  movu                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  pavgb                   m0, [dstq]
+  pavgb                   m1, [dstq+dst_strideq]
+  pavgb                   m2, [dstq+dst_strideq*2]
+  pavgb                   m3, [dstq+r6q]
+%endif
+  mova  [dstq              ], m0
+  mova  [dstq+dst_strideq  ], m1
+  mova  [dstq+dst_strideq*2], m2
+  mova  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop8
+  RET
+
+.w4:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop4:
+  movh                    m0, [srcq]
+  movh                    m1, [srcq+src_strideq]
+  movh                    m2, [srcq+src_strideq*2]
+  movh                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  movh                    m4, [dstq]
+  movh                    m5, [dstq+dst_strideq]
+  movh                    m6, [dstq+dst_strideq*2]
+  movh                    m7, [dstq+r6q]
+  pavgb                   m0, m4
+  pavgb                   m1, m5
+  pavgb                   m2, m6
+  pavgb                   m3, m7
+%endif
+  movh  [dstq              ], m0
+  movh  [dstq+dst_strideq  ], m1
+  movh  [dstq+dst_strideq*2], m2
+  movh  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop4
+  RET
+%endmacro
+
+convolve_fn copy
+convolve_fn avg
diff --git a/media/libvpx/vp9/common/x86/vp9_high_intrapred_sse2.asm b/media/libvpx/vp9/common/x86/vp9_high_intrapred_sse2.asm
new file mode 100644
index 000000000..b12d29c0a
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_high_intrapred_sse2.asm
@@ -0,0 +1,476 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 4 dd 16
+pw_32: times 4 dd 32
+
+SECTION .text
+INIT_MMX sse
+cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  DEFINE_ARGS dst, stride, one
+  mov                 oned, 0x0001
+  pxor                  m1, m1
+  movd                  m3, oned
+  pshufw                m3, m3, 0x0
+  paddw                 m0, m2
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshufw                m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, one
+  mov                 oned, 0x00010001
+  lea             stride3q, [strideq*3]
+  movd                  m3, oned
+  pshufd                m3, m3, 0x0
+  paddw                 m0, m2
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  paddw                 m0, [GLOBAL(pw_8)]
+  psrlw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m3, [aboveq+16]
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_16)]
+  psrad                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova   [dstq              ], m0
+  mova   [dstq           +16], m0
+  mova   [dstq+strideq*2    ], m0
+  mova   [dstq+strideq*2 +16], m0
+  mova   [dstq+strideq*4    ], m0
+  mova   [dstq+strideq*4 +16], m0
+  mova   [dstq+stride3q*2   ], m0
+  mova   [dstq+stride3q*2+16], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+%if ARCH_X86_64
+INIT_XMM sse2
+cglobal highbd_dc_predictor_32x32, 4, 5, 9, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  mova                  m5, [leftq]
+  mova                  m6, [leftq+16]
+  mova                  m7, [leftq+32]
+  mova                  m8, [leftq+48]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  paddw                 m0, m5
+  paddw                 m0, m6
+  paddw                 m0, m7
+  paddw                 m0, m8
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_32)]
+  psrad                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova [dstq               ], m0
+  mova [dstq          +16  ], m0
+  mova [dstq          +32  ], m0
+  mova [dstq          +48  ], m0
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2+16  ], m0
+  mova [dstq+strideq*2+32  ], m0
+  mova [dstq+strideq*2+48  ], m0
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4+16  ], m0
+  mova [dstq+strideq*4+32  ], m0
+  mova [dstq+strideq*4+48  ], m0
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m0
+  mova [dstq+stride3q*2 +32], m0
+  mova [dstq+stride3q*2 +48], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+%endif
+
+INIT_MMX sse
+cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq              ], m0
+  mova    [dstq           +16], m1
+  mova    [dstq+strideq*2    ], m0
+  mova    [dstq+strideq*2 +16], m1
+  mova    [dstq+strideq*4    ], m0
+  mova    [dstq+strideq*4 +16], m1
+  mova    [dstq+stride3q*2   ], m0
+  mova    [dstq+stride3q*2+16], m1
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  mova                  m2, [aboveq+32]
+  mova                  m3, [aboveq+48]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq            +32], m2
+  mova [dstq            +48], m3
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2  +16], m1
+  mova [dstq+strideq*2  +32], m2
+  mova [dstq+strideq*2  +48], m3
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4  +16], m1
+  mova [dstq+strideq*4  +32], m2
+  mova [dstq+strideq*4  +48], m3
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m1
+  mova [dstq+stride3q*2 +32], m2
+  mova [dstq+stride3q*2 +48], m3
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_MMX sse
+cglobal highbd_tm_predictor_4x4, 5, 6, 5, dst, stride, above, left, bps, one
+  movd                  m1, [aboveq-2]
+  movq                  m0, [aboveq]
+  pshufw                m1, m1, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  movd                  m3, oned
+  movd                  m4, bpsd
+  pshufw                m3, m3, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -2
+  mova                  m2, m3
+  psllw                 m3, m4
+  add                leftq, 8
+  psubw                 m3, m2 ; max possible value
+  pxor                  m4, m4 ; min possible value
+  psubw                 m0, m1
+.loop:
+  movq                  m1, [leftq+lineq*4]
+  movq                  m2, [leftq+lineq*4+2]
+  pshufw                m1, m1, 0x0
+  pshufw                m2, m2, 0x0
+  paddw                 m1, m0
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m1, m3
+  pminsw                m2, m3
+  pmaxsw                m1, m4
+  pmaxsw                m2, m4
+  ;Store the values
+  movq    [dstq          ], m1
+  movq    [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_8x8, 5, 6, 5, dst, stride, above, left, bps, one
+  movd                  m1, [aboveq-2]
+  mova                  m0, [aboveq]
+  pshuflw               m1, m1, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  pxor                  m3, m3
+  pxor                  m4, m4
+  pinsrw                m3, oned, 0
+  pinsrw                m4, bpsd, 0
+  pshuflw               m3, m3, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  punpcklqdq            m3, m3
+  mov                lineq, -4
+  mova                  m2, m3
+  punpcklqdq            m1, m1
+  psllw                 m3, m4
+  add                leftq, 16
+  psubw                 m3, m2 ; max possible value
+  pxor                  m4, m4 ; min possible value
+  psubw                 m0, m1
+.loop:
+  movd                  m1, [leftq+lineq*4]
+  movd                  m2, [leftq+lineq*4+2]
+  pshuflw               m1, m1, 0x0
+  pshuflw               m2, m2, 0x0
+  punpcklqdq            m1, m1
+  punpcklqdq            m2, m2
+  paddw                 m1, m0
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m1, m3
+  pminsw                m2, m3
+  pmaxsw                m1, m4
+  pmaxsw                m2, m4
+  ;Store the values
+  mova      [dstq          ], m1
+  mova      [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+%if ARCH_X86_64
+INIT_XMM sse2
+cglobal highbd_tm_predictor_16x16, 5, 6, 9, dst, stride, above, left, bps, one
+  movd                  m2, [aboveq-2]
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  pshuflw               m2, m2, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  pxor                  m7, m7
+  pxor                  m8, m8
+  pinsrw                m7, oned, 0
+  pinsrw                m8, bpsd, 0
+  pshuflw               m7, m7, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  punpcklqdq            m7, m7
+  mov                lineq, -8
+  mova                  m5, m7
+  punpcklqdq            m2, m2
+  psllw                 m7, m8
+  add                leftq, 32
+  psubw                 m7, m5 ; max possible value
+  pxor                  m8, m8 ; min possible value
+  psubw                 m0, m2
+  psubw                 m1, m2
+.loop:
+  movd                  m2, [leftq+lineq*4]
+  movd                  m3, [leftq+lineq*4+2]
+  pshuflw               m2, m2, 0x0
+  pshuflw               m3, m3, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m3, m3
+  paddw                 m4, m2, m0
+  paddw                 m5, m3, m0
+  paddw                 m2, m1
+  paddw                 m3, m1
+  ;Clamp to the bit-depth
+  pminsw                m4, m7
+  pminsw                m5, m7
+  pminsw                m2, m7
+  pminsw                m3, m7
+  pmaxsw                m4, m8
+  pmaxsw                m5, m8
+  pmaxsw                m2, m8
+  pmaxsw                m3, m8
+  ;Store the values
+  mova   [dstq             ], m4
+  mova   [dstq+strideq*2   ], m5
+  mova   [dstq          +16], m2
+  mova   [dstq+strideq*2+16], m3
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_32x32, 5, 6, 12, dst, stride, above, left, bps, one
+  movd                  m0, [aboveq-2]
+  mova                  m1, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  pshuflw               m0, m0, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  pxor                 m10, m10
+  pxor                 m11, m11
+  pinsrw               m10, oned, 0
+  pinsrw               m11, bpsd, 0
+  pshuflw              m10, m10, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  punpcklqdq           m10, m10
+  mov                lineq, -16
+  mova                  m5, m10
+  punpcklqdq            m0, m0
+  psllw                m10, m11
+  add                leftq, 64
+  psubw                m10, m5 ; max possible value
+  pxor                 m11, m11 ; min possible value
+  psubw                 m1, m0
+  psubw                 m2, m0
+  psubw                 m3, m0
+  psubw                 m4, m0
+.loop:
+  movd                  m5, [leftq+lineq*4]
+  movd                  m6, [leftq+lineq*4+2]
+  pshuflw               m5, m5, 0x0
+  pshuflw               m6, m6, 0x0
+  punpcklqdq            m5, m5
+  punpcklqdq            m6, m6
+  paddw                 m7, m5, m1
+  paddw                 m8, m5, m2
+  paddw                 m9, m5, m3
+  paddw                 m5, m4
+  ;Clamp these values to the bit-depth
+  pminsw                m7, m10
+  pminsw                m8, m10
+  pminsw                m9, m10
+  pminsw                m5, m10
+  pmaxsw                m7, m11
+  pmaxsw                m8, m11
+  pmaxsw                m9, m11
+  pmaxsw                m5, m11
+  ;Store these values
+  mova   [dstq           ], m7
+  mova   [dstq        +16], m8
+  mova   [dstq        +32], m9
+  mova   [dstq        +48], m5
+  paddw                 m7, m6, m1
+  paddw                 m8, m6, m2
+  paddw                 m9, m6, m3
+  paddw                 m6, m4
+  ;Clamp these values to the bit-depth
+  pminsw                m7, m10
+  pminsw                m8, m10
+  pminsw                m9, m10
+  pminsw                m6, m10
+  pmaxsw                m7, m11
+  pmaxsw                m8, m11
+  pmaxsw                m9, m11
+  pmaxsw                m6, m11
+  ;Store these values
+  mova   [dstq+strideq*2   ], m7
+  mova   [dstq+strideq*2+16], m8
+  mova   [dstq+strideq*2+32], m9
+  mova   [dstq+strideq*2+48], m6
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+%endif
diff --git a/media/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c b/media/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
new file mode 100644
index 000000000..b40669c63
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
@@ -0,0 +1,1215 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {
+  __m128i ubounded;
+  __m128i lbounded;
+  __m128i retval;
+
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i t80, max, min;
+
+  if (bd == 8) {
+    t80 = _mm_set1_epi16(0x80);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
+  } else if (bd == 10) {
+    t80 = _mm_set1_epi16(0x200);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
+  } else {  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+    max = _mm_subs_epi16(
+              _mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
+  }
+
+  min = _mm_subs_epi16(zero, t80);
+
+  ubounded = _mm_cmpgt_epi16(value, max);
+  lbounded = _mm_cmplt_epi16(value, min);
+  retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
+  ubounded = _mm_and_si128(ubounded, max);
+  lbounded = _mm_and_si128(lbounded, min);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_or_si128(retval, lbounded);
+  return retval;
+}
+
+// TODO(debargha, peter): Break up large functions into smaller ones
+// in this file.
+static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
+                                                   int p,
+                                                   const uint8_t *_blimit,
+                                                   const uint8_t *_limit,
+                                                   const uint8_t *_thresh,
+                                                   int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i blimit, limit, thresh;
+  __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
+  __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
+  __m128i ps1, qs1, ps0, qs0;
+  __m128i abs_p0q0, abs_p1q1, ffff, work;
+  __m128i filt, work_a, filter1, filter2;
+  __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4;
+  __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1;
+  __m128i flat2_q0, flat2_p0;
+  __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0;
+  __m128i pixelFilter_p, pixelFilter_q;
+  __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+  __m128i sum_p7, sum_q7, sum_p3, sum_q3;
+  __m128i t4, t3, t80, t1;
+  __m128i eight, four;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+  }
+
+  q4 = _mm_load_si128((__m128i *)(s + 4 * p));
+  p4 = _mm_load_si128((__m128i *)(s - 5 * p));
+  q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+
+  //  highbd_filter_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+
+  ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+
+  //  highbd_hev_mask (in C code this is actually called from highbd_filter4)
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);  // abs(p0 - q0) * 2
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);  // abs(p1 - q1) / 2
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p1, p0),
+                                    _mm_subs_epu16(p0, p1)),
+                       _mm_or_si128(_mm_subs_epu16(q1, q0),
+                                    _mm_subs_epu16(q0, q1)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);  // return ~mask
+
+  // lp filter
+  // highbd_filter4
+  t4 = _mm_set1_epi16(4);
+  t3 = _mm_set1_epi16(3);
+  if (bd == 8)
+    t80 = _mm_set1_epi16(0x80);
+  else if (bd == 10)
+    t80 = _mm_set1_epi16(0x200);
+  else  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+
+  t1 = _mm_set1_epi16(0x1);
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+
+  filt = _mm_and_si128(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+  filt = _mm_and_si128(filt, mask);
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  filter1 = _mm_srai_epi16(filter1, 0x3);
+  filter2 = _mm_srai_epi16(filter2, 0x3);
+
+  qs0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd),
+      t80);
+  ps0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd),
+      t80);
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  filt = _mm_andnot_si128(hev, filt);
+  qs1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
+      t80);
+  ps1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
+      t80);
+
+  // end highbd_filter4
+  // loopfilter done
+
+  // highbd_flat_mask4
+  flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+                                    _mm_subs_epu16(p0, p2)),
+                       _mm_or_si128(_mm_subs_epu16(p3, p0),
+                                    _mm_subs_epu16(p0, p3)));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q0),
+                                    _mm_subs_epu16(q0, q2)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q0),
+                                    _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  work = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  flat = _mm_max_epi16(work, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  // end flat_mask4
+
+  // flat & mask = flat && mask (as used in filter8)
+  // (because, in both vars, each block of 16 either all 1s or all 0s)
+  flat = _mm_and_si128(flat, mask);
+
+  p5 = _mm_load_si128((__m128i *)(s - 6 * p));
+  q5 = _mm_load_si128((__m128i *)(s + 5 * p));
+  p6 = _mm_load_si128((__m128i *)(s - 7 * p));
+  q6 = _mm_load_si128((__m128i *)(s + 6 * p));
+  p7 = _mm_load_si128((__m128i *)(s - 8 * p));
+  q7 = _mm_load_si128((__m128i *)(s + 7 * p));
+
+  // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7
+  // but referred to as p0-p4 & q0-q4 in fn)
+  flat2 = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p4, p0),
+                                     _mm_subs_epu16(p0, p4)),
+                        _mm_or_si128(_mm_subs_epu16(q4, q0),
+                                     _mm_subs_epu16(q0, q4)));
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p5, p0),
+                                    _mm_subs_epu16(p0, p5)),
+                       _mm_or_si128(_mm_subs_epu16(q5, q0),
+                                    _mm_subs_epu16(q0, q5)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p6, p0),
+                                    _mm_subs_epu16(p0, p6)),
+                       _mm_or_si128(_mm_subs_epu16(q6, q0),
+                                    _mm_subs_epu16(q0, q6)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p7, p0),
+                                    _mm_subs_epu16(p0, p7)),
+                       _mm_or_si128(_mm_subs_epu16(q7, q0),
+                                    _mm_subs_epu16(q0, q7)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  if (bd == 8)
+    flat2 = _mm_subs_epu16(flat2, one);
+  else if (bd == 10)
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
+
+  flat2 = _mm_cmpeq_epi16(flat2, zero);
+  flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  // end highbd_flat_mask5
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // flat and wide flat calculations
+  eight = _mm_set1_epi16(8);
+  four = _mm_set1_epi16(4);
+
+  pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5),
+                                _mm_add_epi16(p4, p3));
+  pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5),
+                                _mm_add_epi16(q4, q3));
+
+  pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1));
+  pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+  pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1));
+  pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+  pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+                                                      pixelFilter_q));
+  pixetFilter_p2p1p0 =   _mm_add_epi16(four,
+                                       _mm_add_epi16(pixetFilter_p2p1p0,
+                                                     pixetFilter_q2q1q0));
+  flat2_p0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(p7, p0)), 4);
+  flat2_q0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(q7, q0)), 4);
+  flat_p0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(p3, p0)), 3);
+  flat_q0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(q3, q0)), 3);
+
+  sum_p7 = _mm_add_epi16(p7, p7);
+  sum_q7 = _mm_add_epi16(q7, q7);
+  sum_p3 = _mm_add_epi16(p3, p3);
+  sum_q3 = _mm_add_epi16(q3, q3);
+
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6);
+  flat2_p1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4);
+  flat2_q1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4);
+
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2);
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2);
+  flat_p1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(sum_p3, p1)), 3);
+  flat_q1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                         _mm_add_epi16(sum_q3, q1)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  sum_p3 = _mm_add_epi16(sum_p3, p3);
+  sum_q3 = _mm_add_epi16(sum_q3, q3);
+
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5);
+  flat2_p2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p2)), 4);
+  flat2_q2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q2)), 4);
+
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1);
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1);
+  flat_p2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                         _mm_add_epi16(sum_p3, p2)), 3);
+  flat_q2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                         _mm_add_epi16(sum_q3, q2)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4);
+  flat2_p3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p3)), 4);
+  flat2_q3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q3)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3);
+  flat2_p4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p4)), 4);
+  flat2_q4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q4)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2);
+  flat2_p5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p5)), 4);
+  flat2_q5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q5)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1);
+  flat2_p6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                          _mm_add_epi16(sum_p7, p6)), 4);
+  flat2_q6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                                          _mm_add_epi16(sum_q7, q6)), 4);
+
+  //  wide flat
+  //  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+  //  highbd_filter8
+  p2 = _mm_andnot_si128(flat, p2);
+  //  p2 remains unchanged if !(flat && mask)
+  flat_p2 = _mm_and_si128(flat, flat_p2);
+  //  when (flat && mask)
+  p2 = _mm_or_si128(p2, flat_p2);  // full list of p2 values
+  q2 = _mm_andnot_si128(flat, q2);
+  flat_q2 = _mm_and_si128(flat, flat_q2);
+  q2 = _mm_or_si128(q2, flat_q2);  // full list of q2 values
+
+  ps1 = _mm_andnot_si128(flat, ps1);
+  //  p1 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p1 = _mm_and_si128(flat, flat_p1);
+  //  when (flat && mask)
+  p1 = _mm_or_si128(ps1, flat_p1);  // full list of p1 values
+  qs1 = _mm_andnot_si128(flat, qs1);
+  flat_q1 = _mm_and_si128(flat, flat_q1);
+  q1 = _mm_or_si128(qs1, flat_q1);  // full list of q1 values
+
+  ps0 = _mm_andnot_si128(flat, ps0);
+  //  p0 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p0 = _mm_and_si128(flat, flat_p0);
+  //  when (flat && mask)
+  p0 = _mm_or_si128(ps0, flat_p0);  // full list of p0 values
+  qs0 = _mm_andnot_si128(flat, qs0);
+  flat_q0 = _mm_and_si128(flat, flat_q0);
+  q0 = _mm_or_si128(qs0, flat_q0);  // full list of q0 values
+  // end highbd_filter8
+
+  // highbd_filter16
+  p6 = _mm_andnot_si128(flat2, p6);
+  //  p6 remains unchanged if !(flat2 && flat && mask)
+  flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+  //  get values for when (flat2 && flat && mask)
+  p6 = _mm_or_si128(p6, flat2_p6);  // full list of p6 values
+  q6 = _mm_andnot_si128(flat2, q6);
+  //  q6 remains unchanged if !(flat2 && flat && mask)
+  flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+  //  get values for when (flat2 && flat && mask)
+  q6 = _mm_or_si128(q6, flat2_q6);  // full list of q6 values
+  _mm_store_si128((__m128i *)(s - 7 * p), p6);
+  _mm_store_si128((__m128i *)(s + 6 * p), q6);
+
+  p5 = _mm_andnot_si128(flat2, p5);
+  //  p5 remains unchanged if !(flat2 && flat && mask)
+  flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+  //  get values for when (flat2 && flat && mask)
+  p5 = _mm_or_si128(p5, flat2_p5);
+  //  full list of p5 values
+  q5 = _mm_andnot_si128(flat2, q5);
+  //  q5 remains unchanged if !(flat2 && flat && mask)
+  flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+  //  get values for when (flat2 && flat && mask)
+  q5 = _mm_or_si128(q5, flat2_q5);
+  //  full list of q5 values
+  _mm_store_si128((__m128i *)(s - 6 * p), p5);
+  _mm_store_si128((__m128i *)(s + 5 * p), q5);
+
+  p4 = _mm_andnot_si128(flat2, p4);
+  //  p4 remains unchanged if !(flat2 && flat && mask)
+  flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+  //  get values for when (flat2 && flat && mask)
+  p4 = _mm_or_si128(p4, flat2_p4);  // full list of p4 values
+  q4 = _mm_andnot_si128(flat2, q4);
+  //  q4 remains unchanged if !(flat2 && flat && mask)
+  flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+  //  get values for when (flat2 && flat && mask)
+  q4 = _mm_or_si128(q4, flat2_q4);  // full list of q4 values
+  _mm_store_si128((__m128i *)(s - 5 * p), p4);
+  _mm_store_si128((__m128i *)(s + 4 * p), q4);
+
+  p3 = _mm_andnot_si128(flat2, p3);
+  //  p3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+  //  get values for when (flat2 && flat && mask)
+  p3 = _mm_or_si128(p3, flat2_p3);  // full list of p3 values
+  q3 = _mm_andnot_si128(flat2, q3);
+  //  q3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+  //  get values for when (flat2 && flat && mask)
+  q3 = _mm_or_si128(q3, flat2_q3);  // full list of q3 values
+  _mm_store_si128((__m128i *)(s - 4 * p), p3);
+  _mm_store_si128((__m128i *)(s + 3 * p), q3);
+
+  p2 = _mm_andnot_si128(flat2, p2);
+  //  p2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+  //  get values for when (flat2 && flat && mask)
+  p2 = _mm_or_si128(p2, flat2_p2);
+  //  full list of p2 values
+  q2 = _mm_andnot_si128(flat2, q2);
+  //  q2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+  //  get values for when (flat2 && flat && mask)
+  q2 = _mm_or_si128(q2, flat2_q2);  // full list of q2 values
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+
+  p1 = _mm_andnot_si128(flat2, p1);
+  //  p1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+  //  get values for when (flat2 && flat && mask)
+  p1 = _mm_or_si128(p1, flat2_p1);  // full list of p1 values
+  q1 = _mm_andnot_si128(flat2, q1);
+  //  q1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+  //  get values for when (flat2 && flat && mask)
+  q1 = _mm_or_si128(q1, flat2_q1);  // full list of q1 values
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+
+  p0 = _mm_andnot_si128(flat2, p0);
+  //  p0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+  //  get values for when (flat2 && flat && mask)
+  p0 = _mm_or_si128(p0, flat2_p0);  // full list of p0 values
+  q0 = _mm_andnot_si128(flat2, q0);
+  //  q0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+  //  get values for when (flat2 && flat && mask)
+  q0 = _mm_or_si128(q0, flat2_q0);  // full list of q0 values
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s - 0 * p), q0);
+}
+
+static void highbd_mb_lpf_horizontal_edge_w_sse2_16(uint16_t *s,
+                                                    int p,
+                                                    const uint8_t *_blimit,
+                                                    const uint8_t *_limit,
+                                                    const uint8_t *_thresh,
+                                                    int bd) {
+  highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
+  highbd_mb_lpf_horizontal_edge_w_sse2_8(s + 8, p, _blimit, _limit, _thresh,
+                                         bd);
+}
+
+// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
+void vp9_highbd_lpf_horizontal_16_sse2(uint16_t *s, int p,
+                                       const uint8_t *_blimit,
+                                       const uint8_t *_limit,
+                                       const uint8_t *_thresh,
+                                       int count, int bd) {
+  if (count == 1)
+    highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
+  else
+    highbd_mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh, bd);
+}
+
+void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh,
+                                      int count, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, flat_op2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op0[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i ffff = _mm_cmpeq_epi16(one, one);
+  __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+  const __m128i four = _mm_set1_epi16(4);
+  __m128i workp_a, workp_b, workp_shft;
+
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  (void)count;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_set1_epi16(0x200);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+          _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_set1_epi16(0x800);
+  }
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+
+  // filter_mask and hev_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+                          _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+                          _mm_subs_epu16(q0, q1));
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+                          _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+                          _mm_subs_epu16(q1, p1));
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(abs_p1p0, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  mask = _mm_max_epi16(abs_q1q0, mask);
+  // mask |= (abs(q1 - q0) > limit) * -1;
+
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // flat_mask4
+  flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+                                    _mm_subs_epu16(p0, p2)),
+                       _mm_or_si128(_mm_subs_epu16(q2, q0),
+                                    _mm_subs_epu16(q0, q2)));
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p0),
+                                    _mm_subs_epu16(p0, p3)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q0),
+                                    _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  flat = _mm_max_epi16(abs_p1p0, flat);
+  flat = _mm_max_epi16(abs_q1q0, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);  // flat & mask
+
+  // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+  workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+  workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op2[0], workp_shft);
+
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft);
+
+  // lp filter
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  // (vp9_filter + 3 * (qs0 - ps0)) & mask
+  filt = signed_char_clamp_bd_sse2(filt, bd);
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = _mm_adds_epi16(filt, t4);
+  filter2 = _mm_adds_epi16(filt, t3);
+
+  // Filter1 >> 3
+  filter1 = signed_char_clamp_bd_sse2(filter1, bd);
+  filter1 = _mm_srai_epi16(filter1, 3);
+
+  // Filter2 >> 3
+  filter2 = signed_char_clamp_bd_sse2(filter2, bd);
+  filter2 = _mm_srai_epi16(filter2, 3);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+  filt = _mm_andnot_si128(hev, filt);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q0 = _mm_load_si128((__m128i *)flat_oq0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q0 = _mm_and_si128(flat, q0);
+  q0 = _mm_or_si128(work_a, q0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q1 = _mm_load_si128((__m128i *)flat_oq1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q1 = _mm_and_si128(flat, q1);
+  q1 = _mm_or_si128(work_a, q1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q2 = _mm_load_si128((__m128i *)flat_oq2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q2 = _mm_and_si128(flat, q2);
+  q2 = _mm_or_si128(work_a, q2);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p0 = _mm_load_si128((__m128i *)flat_op0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p0 = _mm_and_si128(flat, p0);
+  p0 = _mm_or_si128(work_a, p0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p1 = _mm_load_si128((__m128i *)flat_op1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p1 = _mm_and_si128(flat, p1);
+  p1 = _mm_or_si128(work_a, p1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p2 = _mm_load_si128((__m128i *)flat_op2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p2 = _mm_and_si128(flat, p2);
+  p2 = _mm_or_si128(work_a, p2);
+
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s + 0 * p), q0);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+}
+
+void vp9_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit0,
+                                           const uint8_t *_limit0,
+                                           const uint8_t *_thresh0,
+                                           const uint8_t *_blimit1,
+                                           const uint8_t *_limit1,
+                                           const uint8_t *_thresh1,
+                                           int bd) {
+  vp9_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1,
+                                   1, bd);
+}
+
+void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh,
+                                      int count, int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+                                        _mm_subs_epu16(p0, p1));
+  const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+                                        _mm_subs_epu16(q0, q1));
+  const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+                                  _mm_subs_epu16(q0, p0));
+  __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+                                  _mm_subs_epu16(q1, p1));
+  __m128i work;
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  __m128i tff80;
+  __m128i tffe0;
+  __m128i t1f;
+  // equivalent to shifting 0x1f left by bitdepth - 8
+  // and setting new bits to 1
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i t7f;
+  // equivalent to shifting 0x7f left by bitdepth - 8
+  // and setting new bits to 1
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  (void)count;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+    tff80 = _mm_set1_epi16(0xff80);
+    tffe0 = _mm_set1_epi16(0xffe0);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
+  }
+
+  ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+  ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+  qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+  qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+
+  // filter_mask and hev_mask
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(flat, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  // mask |= (abs(q1 - q0) > limit) * -1;
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+                                    _mm_subs_epu16(p1, p2)),
+                       _mm_or_si128(_mm_subs_epu16(p3, p2),
+                                    _mm_subs_epu16(p2, p3)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q1),
+                                    _mm_subs_epu16(q1, q2)),
+                       _mm_or_si128(_mm_subs_epu16(q3, q2),
+                                    _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // filter4
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+
+  // (vp9_filter + 3 * (qs0 - ps0)) & mask
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter1);  // get the values that are <0
+  filter1 = _mm_srli_epi16(filter1, 3);
+  work_a = _mm_and_si128(work_a, tffe0);  // sign bits for the values < 0
+  filter1 = _mm_and_si128(filter1, t1f);  // clamp the range
+  filter1 = _mm_or_si128(filter1, work_a);  // reinsert the sign bits
+
+  // Filter2 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter2);
+  filter2 = _mm_srli_epi16(filter2, 3);
+  work_a = _mm_and_si128(work_a, tffe0);
+  filter2 = _mm_and_si128(filter2, t1f);
+  filter2 = _mm_or_si128(filter2, work_a);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  work_a = _mm_cmpgt_epi16(zero, filt);
+  filt = _mm_srli_epi16(filt, 1);
+  work_a = _mm_and_si128(work_a, tff80);
+  filt = _mm_and_si128(filt, t7f);
+  filt = _mm_or_si128(filt, work_a);
+
+  filt = _mm_andnot_si128(hev, filt);
+
+  q0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
+  q1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), t80);
+  p0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
+  p1 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), t80);
+
+  _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+}
+
+void vp9_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit0,
+                                           const uint8_t *_limit0,
+                                           const uint8_t *_thresh0,
+                                           const uint8_t *_blimit1,
+                                           const uint8_t *_limit1,
+                                           const uint8_t *_thresh1,
+                                           int bd) {
+  vp9_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
+  vp9_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, 1,
+                                   bd);
+}
+
+static INLINE void highbd_transpose(uint16_t *src[], int in_p,
+                                    uint16_t *dst[], int out_p,
+                                    int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    uint16_t *in = src[idx8x8];
+    uint16_t *out = dst[idx8x8];
+
+    p0 = _mm_loadu_si128((__m128i *)(in + 0*in_p));  // 00 01 02 03 04 05 06 07
+    p1 = _mm_loadu_si128((__m128i *)(in + 1*in_p));  // 10 11 12 13 14 15 16 17
+    p2 = _mm_loadu_si128((__m128i *)(in + 2*in_p));  // 20 21 22 23 24 25 26 27
+    p3 = _mm_loadu_si128((__m128i *)(in + 3*in_p));  // 30 31 32 33 34 35 36 37
+    p4 = _mm_loadu_si128((__m128i *)(in + 4*in_p));  // 40 41 42 43 44 45 46 47
+    p5 = _mm_loadu_si128((__m128i *)(in + 5*in_p));  // 50 51 52 53 54 55 56 57
+    p6 = _mm_loadu_si128((__m128i *)(in + 6*in_p));  // 60 61 62 63 64 65 66 67
+    p7 = _mm_loadu_si128((__m128i *)(in + 7*in_p));  // 70 71 72 73 74 75 76 77
+    // 00 10 01 11 02 12 03 13
+    x0 = _mm_unpacklo_epi16(p0, p1);
+    // 20 30 21 31 22 32 23 33
+    x1 = _mm_unpacklo_epi16(p2, p3);
+    // 40 50 41 51 42 52 43 53
+    x2 = _mm_unpacklo_epi16(p4, p5);
+    // 60 70 61 71 62 72 63 73
+    x3 = _mm_unpacklo_epi16(p6, p7);
+    // 00 10 20 30 01 11 21 31
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 40 50 60 70 41 51 61 71
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 00 10 20 30 40 50 60 70
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 01 11 21 31 41 51 61 71
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 0*out_p), x6);
+    // 00 10 20 30 40 50 60 70
+    _mm_storeu_si128((__m128i *)(out + 1*out_p), x7);
+    // 01 11 21 31 41 51 61 71
+
+    // 02 12 22 32 03 13 23 33
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 42 52 62 72 43 53 63 73
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 02 12 22 32 42 52 62 72
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 2*out_p), x6);
+    // 02 12 22 32 42 52 62 72
+    _mm_storeu_si128((__m128i *)(out + 3*out_p), x7);
+    // 03 13 23 33 43 53 63 73
+
+    // 04 14 05 15 06 16 07 17
+    x0 = _mm_unpackhi_epi16(p0, p1);
+    // 24 34 25 35 26 36 27 37
+    x1 = _mm_unpackhi_epi16(p2, p3);
+    // 44 54 45 55 46 56 47 57
+    x2 = _mm_unpackhi_epi16(p4, p5);
+    // 64 74 65 75 66 76 67 77
+    x3 = _mm_unpackhi_epi16(p6, p7);
+    // 04 14 24 34 05 15 25 35
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 44 54 64 74 45 55 65 75
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 04 14 24 34 44 54 64 74
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 05 15 25 35 45 55 65 75
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 4*out_p), x6);
+    // 04 14 24 34 44 54 64 74
+    _mm_storeu_si128((__m128i *)(out + 5*out_p), x7);
+    // 05 15 25 35 45 55 65 75
+
+    // 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 06 16 26 36 46 56 66 76
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 6*out_p), x6);
+    // 06 16 26 36 46 56 66 76
+    _mm_storeu_si128((__m128i *)(out + 7*out_p), x7);
+    // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1,
+                                        int in_p, uint16_t *out, int out_p) {
+  uint16_t *src0[1];
+  uint16_t *src1[1];
+  uint16_t *dest0[1];
+  uint16_t *dest1[1];
+  src0[0] = in0;
+  src1[0] = in1;
+  dest0[0] = out;
+  dest1[0] = out + 8;
+  highbd_transpose(src0, in_p, dest0, out_p, 1);
+  highbd_transpose(src1, in_p, dest1, out_p, 1);
+}
+
+void vp9_highbd_lpf_vertical_4_sse2(uint16_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh,
+                                    int count, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+  (void)count;
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vp9_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
+                                   bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vp9_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int p,
+                                         const uint8_t *blimit0,
+                                         const uint8_t *limit0,
+                                         const uint8_t *thresh0,
+                                         const uint8_t *blimit1,
+                                         const uint8_t *limit1,
+                                         const uint8_t *thresh1,
+                                         int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vp9_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_8_sse2(uint16_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh,
+                                    int count, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+  (void)count;
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vp9_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
+                                   bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vp9_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int p,
+                                         const uint8_t *blimit0,
+                                         const uint8_t *limit0,
+                                         const uint8_t *thresh0,
+                                         const uint8_t *blimit1,
+                                         const uint8_t *limit1,
+                                         const uint8_t *thresh1,
+                                         int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vp9_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_16_sse2(uint16_t *s, int p,
+                                     const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh,
+                                     int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  highbd_transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  highbd_mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit,
+                                         thresh, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_16_dual_sse2(uint16_t *s,
+                                          int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh,
+                                          int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[256]);
+
+  //  Transpose 16x16
+  highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  //  Loop filtering
+  highbd_mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
+                                          thresh, bd);
+
+  //  Transpose back
+  highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/media/libvpx/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm b/media/libvpx/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm
new file mode 100644
index 000000000..29ec151ed
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm
@@ -0,0 +1,962 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro HIGH_GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm6
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+    punpcklwd   xmm1, xmm7
+
+    movdqa      k0k6, xmm0
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+    movdqa      k1k7, xmm1
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)      ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+    punpcklwd   xmm0, xmm6                  ;two row in one register
+    punpcklwd   xmm1, xmm7
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+
+    pmaddwd     xmm0, k0k6                  ;multiply the filter factors
+    pmaddwd     xmm1, k1k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm3, k3k4
+
+    paddd       xmm0, xmm1                  ;sum
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+
+    paddd       xmm0, krd                   ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+%endm
+
+%macro HIGH_GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm6, xmm7
+    punpckhwd   xmm2, xmm5
+    punpckhwd   xmm3, xmm4
+
+    movdqa      k0k1, xmm0                  ;store filter factors on stack
+    movdqa      k6k7, xmm6
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+%endm
+
+%macro LOAD_VERT_8 1
+    movdqu      xmm0, [rsi + %1]            ;0
+    movdqu      xmm1, [rsi + rax + %1]      ;1
+    movdqu      xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movdqu      xmm7, [rsi + rdx * 2 + %1]  ;7
+    movdqu      xmm2, [rsi + rax + %1]      ;2
+    movdqu      xmm3, [rsi + rax * 2 + %1]  ;3
+    movdqu      xmm4, [rsi + rdx + %1]      ;4
+    movdqu      xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro HIGH_APPLY_FILTER_8 2
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm4, xmm1
+    movdqa      xmm1, xmm6
+    punpcklwd   xmm6, xmm7
+    punpckhwd   xmm1, xmm7
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm2, xmm5
+    punpckhwd   xmm7, xmm5
+
+    movdqu      xmm5, temp
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm3
+    punpcklwd   xmm3, xmm5
+    punpckhwd   xmm4, xmm5
+    movdqu      xmm5, temp
+
+    pmaddwd     xmm0, k0k1
+    pmaddwd     xmm5, k0k1
+    pmaddwd     xmm6, k6k7
+    pmaddwd     xmm1, k6k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm7, k2k5
+    pmaddwd     xmm3, k3k4
+    pmaddwd     xmm4, k3k4
+
+    paddd       xmm0, xmm6
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+    paddd       xmm5, xmm1
+    paddd       xmm5, xmm7
+    paddd       xmm5, xmm4
+
+    paddd       xmm0, krd                   ;rounding
+    paddd       xmm5, krd
+    psrad       xmm0, 7                     ;shift
+    psrad       xmm5, 7
+    packssdw    xmm0, xmm5                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movdqu      xmm1, [rdi + %2]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi + %2], xmm0
+%endm
+
+;void vp9_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d4_v8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d8_v8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d16_v8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 0, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 1, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d4_h8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d8_h8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_highbd_filter_block1d16_h8_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 0, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 1, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/media/libvpx/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm b/media/libvpx/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm
new file mode 100644
index 000000000..93784121c
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm
@@ -0,0 +1,494 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro HIGH_GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklwd   xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm5, rdx
+    movq        xmm2, rcx
+    pshufd      xmm5, xmm5, 0b
+    movdqa      xmm1, xmm5
+    psllw       xmm5, xmm2
+    psubw       xmm5, xmm1                  ;max value (for clamping)
+    pxor        xmm2, xmm2                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+
+    punpcklwd   xmm0, xmm1                  ;two row in one register
+    pmaddwd     xmm0, xmm4                  ;multiply the filter factors
+
+    paddd       xmm0, xmm3                  ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm5
+    pmaxsw      xmm0, xmm2
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+
+    movq        [rdi], xmm0
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%if ARCH_X86_64
+%macro HIGH_GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm6, [rdx]                 ;load filters
+
+    pshuflw     xmm7, xmm6, 11111111b       ;k3
+    pshufhw     xmm6, xmm6, 0b              ;k4
+    psrldq      xmm6, 8
+    punpcklwd   xmm7, xmm6                  ;k3k4k3k4k3k4k3k4
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm8, rdx
+    movq        xmm5, rcx
+    pshufd      xmm8, xmm8, 0b
+    movdqa      xmm1, xmm8
+    psllw       xmm8, xmm5
+    psubw       xmm8, xmm1                  ;max value (for clamping)
+    pxor        xmm5, xmm5                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_8 1
+    movdqa      xmm6, xmm0
+    punpckhwd   xmm6, xmm1
+    punpcklwd   xmm0, xmm1
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+
+    paddd       xmm6, xmm4                  ;rounding
+    paddd       xmm0, xmm4                  ;rounding
+    psrad       xmm6, 7                     ;shift
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%macro HIGH_APPLY_FILTER_16 1
+    movdqa      xmm9, xmm0
+    movdqa      xmm6, xmm2
+    punpckhwd   xmm9, xmm1
+    punpckhwd   xmm6, xmm3
+    punpcklwd   xmm0, xmm1
+    punpcklwd   xmm2, xmm3
+
+    pmaddwd     xmm9, xmm7
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+    pmaddwd     xmm2, xmm7
+
+    paddd       xmm9, xmm4                  ;rounding
+    paddd       xmm6, xmm4
+    paddd       xmm0, xmm4
+    paddd       xmm2, xmm4
+
+    psrad       xmm9, 7                     ;shift
+    psrad       xmm6, 7
+    psrad       xmm0, 7
+    psrad       xmm2, 7
+
+    packssdw    xmm0, xmm9                  ;pack back to word
+    packssdw    xmm2, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+    pminsw      xmm2, xmm8
+    pmaxsw      xmm2, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    movdqu      xmm3, [rdi + 16]
+    pavgw       xmm0, xmm1
+    pavgw       xmm2, xmm3
+%endif
+    movdqu      [rdi], xmm0               ;store the result
+    movdqu      [rdi + 16], xmm2          ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+%endif
+
+global sym(vp9_highbd_filter_block1d4_v2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vp9_highbd_filter_block1d8_v2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_v2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vp9_highbd_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vp9_highbd_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vp9_highbd_filter_block1d4_h2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vp9_highbd_filter_block1d8_h2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_h2_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(vp9_highbd_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(vp9_highbd_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_highbd_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(vp9_highbd_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
diff --git a/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c b/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
new file mode 100644
index 000000000..ce010df3b
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -0,0 +1,4223 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/x86/vp9_idct_intrin_sse2.h"
+#include "vp9/common/vp9_idct.h"
+
+#define RECON_AND_STORE4X4(dest, in_x) \
+{                                                     \
+  __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
+  d0 = _mm_unpacklo_epi8(d0, zero); \
+  d0 = _mm_add_epi16(in_x, d0); \
+  d0 = _mm_packus_epi16(d0, d0); \
+  *(int *)(dest) = _mm_cvtsi128_si32(d0); \
+}
+
+void vp9_idct4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i cst = _mm_setr_epi16(
+      (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
+      (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
+      (int16_t)cospi_8_64, (int16_t)cospi_24_64);
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i input0, input1, input2, input3;
+
+  // Rows
+  input0 = _mm_load_si128((const __m128i *)input);
+  input2 = _mm_load_si128((const __m128i *)(input + 8));
+
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_shufflelo_epi16(input0, 0xd8);
+  input0 = _mm_shufflehi_epi16(input0, 0xd8);
+  input2 = _mm_shufflelo_epi16(input2, 0xd8);
+  input2 = _mm_shufflehi_epi16(input2, 0xd8);
+
+  input1 = _mm_unpackhi_epi32(input0, input0);
+  input0 = _mm_unpacklo_epi32(input0, input0);
+  input3 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpacklo_epi32(input2, input2);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input1);
+  input1 = _mm_packs_epi32(input2, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Columns
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_unpacklo_epi32(input2, input2);
+  input1 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpackhi_epi32(input3, input3);
+  input3 = _mm_unpacklo_epi32(input3, input3);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input2);
+  input1 = _mm_packs_epi32(input1, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Final round and shift
+  input2 = _mm_add_epi16(input2, eight);
+  input3 = _mm_add_epi16(input3, eight);
+
+  input2 = _mm_srai_epi16(input2, 4);
+  input3 = _mm_srai_epi16(input3, 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, input2);
+    d2 = _mm_add_epi16(d2, input3);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store input0
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store input1
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store input2
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+    // store input3
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void vp9_idct4x4_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 4);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
+}
+
+static void idct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+
+  transpose_4x4(in);
+  // stage 1
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[3], v[2]);
+
+  // stage 2
+  in[0] = _mm_add_epi16(u[0], u[1]);
+  in[1] = _mm_sub_epi16(u[0], u[1]);
+  in[1] = _mm_shuffle_epi32(in[1], 0x4E);
+}
+
+static void iadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
+  const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8], in7;
+
+  transpose_4x4(in);
+  in7 = _mm_srli_si128(in[1], 8);
+  in7 = _mm_add_epi16(in7, in[0]);
+  in7 = _mm_sub_epi16(in7, in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpackhi_epi16(in[0], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(v[3], v[4]);
+  u[2] = v[2];
+  u[3] = _mm_add_epi32(u[0], u[1]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_add_epi32(u[3], v[5]);
+  u[6] = _mm_sub_epi32(u[5], u[4]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+}
+
+void vp9_iht4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[2];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+
+  in[0] = _mm_loadu_si128((const __m128i *)(input));
+  in[1] = _mm_loadu_si128((const __m128i *)(input + 8));
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct4_sse2(in);
+      idct4_sse2(in);
+      break;
+    case 1:  // ADST_DCT
+      idct4_sse2(in);
+      iadst4_sse2(in);
+      break;
+    case 2:  // DCT_ADST
+      iadst4_sse2(in);
+      idct4_sse2(in);
+      break;
+    case 3:  // ADST_ADST
+      iadst4_sse2(in);
+      iadst4_sse2(in);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  // Final round and shift
+  in[0] = _mm_add_epi16(in[0], eight);
+  in[1] = _mm_add_epi16(in[1], eight);
+
+  in[0] = _mm_srai_epi16(in[0], 4);
+  in[1] = _mm_srai_epi16(in[1], 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        d2, _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)));
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, in[0]);
+    d2 = _mm_add_epi16(d2, in[1]);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store result[0]
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store result[1]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store result[2]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+    // store result[3]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, \
+                      out0, out1, out2, out3, out4, out5, out6, out7) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1); \
+    const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5); \
+    const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7); \
+    const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5); \
+    const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7); \
+                                                        \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7); \
+                                                            \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+    out4 = _mm_unpacklo_epi64(tr1_1, tr1_5); \
+    out5 = _mm_unpackhi_epi64(tr1_1, tr1_5); \
+    out6 = _mm_unpacklo_epi64(tr1_3, tr1_7); \
+    out7 = _mm_unpackhi_epi64(tr1_3, tr1_7); \
+  }
+
+#define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, \
+                         out0, out1, out2, out3) \
+  {                                              \
+    const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0); \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3); \
+    const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2); \
+    \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5); \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5); \
+    \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4); \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4); \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6); \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6); \
+  }
+
+#define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
+  {                                            \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+    out0 = _mm_unpacklo_epi32(tr0_0, tr0_1); \
+    out1 = _mm_unpackhi_epi32(tr0_0, tr0_1); \
+  }
+
+// Define Macro for multiplying elements by constants and adding them together.
+#define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, \
+                               cst0, cst1, cst2, cst3, res0, res1, res2, res3) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      tmp4 = _mm_madd_epi16(lo_1, cst2); \
+      tmp5 = _mm_madd_epi16(hi_1, cst2); \
+      tmp6 = _mm_madd_epi16(lo_1, cst3); \
+      tmp7 = _mm_madd_epi16(hi_1, cst3); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      tmp4 = _mm_add_epi32(tmp4, rounding); \
+      tmp5 = _mm_add_epi32(tmp5, rounding); \
+      tmp6 = _mm_add_epi32(tmp6, rounding); \
+      tmp7 = _mm_add_epi32(tmp7, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS); \
+      tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS); \
+      tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS); \
+      tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+      res2 = _mm_packs_epi32(tmp4, tmp5); \
+      res3 = _mm_packs_epi32(tmp6, tmp7); \
+  }
+
+#define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
+  {   \
+      tmp0 = _mm_madd_epi16(lo_0, cst0); \
+      tmp1 = _mm_madd_epi16(hi_0, cst0); \
+      tmp2 = _mm_madd_epi16(lo_0, cst1); \
+      tmp3 = _mm_madd_epi16(hi_0, cst1); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      res0 = _mm_packs_epi32(tmp0, tmp1); \
+      res1 = _mm_packs_epi32(tmp2, tmp3); \
+  }
+
+#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, \
+                 out0, out1, out2, out3, out4, out5, out6, out7)  \
+  { \
+  /* Stage1 */      \
+  { \
+    const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7); \
+    const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7); \
+    const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5); \
+    const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5); \
+    \
+    MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, \
+                          stg1_1, stg1_2, stg1_3, stp1_4,      \
+                          stp1_7, stp1_5, stp1_6)              \
+  } \
+    \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4); \
+    const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4); \
+    const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6); \
+    const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, \
+                           stg2_1, stg2_2, stg2_3, stp2_0,     \
+                           stp2_1, stp2_2, stp2_3)             \
+    \
+    stp2_4 = _mm_adds_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_subs_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_subs_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_adds_epi16(stp1_7, stp1_6); \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_adds_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_adds_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_subs_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_subs_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_56, stg2_1); \
+    tmp1 = _mm_madd_epi16(hi_56, stg2_1); \
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0); \
+    tmp3 = _mm_madd_epi16(hi_56, stg2_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  } \
+  \
+  /* Stage4  */ \
+  out0 = _mm_adds_epi16(stp1_0, stp2_7); \
+  out1 = _mm_adds_epi16(stp1_1, stp1_6); \
+  out2 = _mm_adds_epi16(stp1_2, stp1_5); \
+  out3 = _mm_adds_epi16(stp1_3, stp2_4); \
+  out4 = _mm_subs_epi16(stp1_3, stp2_4); \
+  out5 = _mm_subs_epi16(stp1_2, stp1_5); \
+  out6 = _mm_subs_epi16(stp1_1, stp1_6); \
+  out7 = _mm_subs_epi16(stp1_0, stp2_7); \
+  }
+
+void vp9_idct8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data.
+  in0 = _mm_load_si128((const __m128i *)input);
+  in1 = _mm_load_si128((const __m128i *)(input + 8 * 1));
+  in2 = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in3 = _mm_load_si128((const __m128i *)(input + 8 * 3));
+  in4 = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in5 = _mm_load_si128((const __m128i *)(input + 8 * 5));
+  in6 = _mm_load_si128((const __m128i *)(input + 8 * 6));
+  in7 = _mm_load_si128((const __m128i *)(input + 8 * 7));
+
+  // 2-D
+  for (i = 0; i < 2; i++) {
+    // 8x8 Transpose is copied from vp9_fdct8x8_sse2()
+    TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7,
+                  in0, in1, in2, in3, in4, in5, in6, in7);
+
+    // 4-stage 1D idct8x8
+    IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+          in0, in1, in2, in3, in4, in5, in6, in7);
+  }
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+void vp9_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 5);
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE(dest + 0 * stride, dc_value);
+  RECON_AND_STORE(dest + 1 * stride, dc_value);
+  RECON_AND_STORE(dest + 2 * stride, dc_value);
+  RECON_AND_STORE(dest + 3 * stride, dc_value);
+  RECON_AND_STORE(dest + 4 * stride, dc_value);
+  RECON_AND_STORE(dest + 5 * stride, dc_value);
+  RECON_AND_STORE(dest + 6 * stride, dc_value);
+  RECON_AND_STORE(dest + 7 * stride, dc_value);
+}
+
+static void idct8_sse2(__m128i *in) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // 8x8 Transpose is copied from vp9_fdct8x8_sse2()
+  TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7],
+                in0, in1, in2, in3, in4, in5, in6, in7);
+
+  // 4-stage 1D idct8x8
+  IDCT8(in0, in1, in2, in3, in4, in5, in6, in7,
+        in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7]);
+}
+
+static void iadst8_sse2(__m128i *in) {
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // transpose
+  array_transpose_8x8(in, in);
+
+  // properly aligned for butterfly input
+  in0 = in[7];
+  in1 = in[0];
+  in2 = in[5];
+  in3 = in[2];
+  in4 = in[3];
+  in5 = in[4];
+  in6 = in[1];
+  in7 = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+}
+
+void vp9_iht8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[8];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+
+  // load input data
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 8 * 1));
+  in[2] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in[3] = _mm_load_si128((const __m128i *)(input + 8 * 3));
+  in[4] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in[5] = _mm_load_si128((const __m128i *)(input + 8 * 5));
+  in[6] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+  in[7] = _mm_load_si128((const __m128i *)(input + 8 * 7));
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct8_sse2(in);
+      idct8_sse2(in);
+      break;
+    case 1:  // ADST_DCT
+      idct8_sse2(in);
+      iadst8_sse2(in);
+      break;
+    case 2:  // DCT_ADST
+      iadst8_sse2(in);
+      idct8_sse2(in);
+      break;
+    case 3:  // ADST_ADST
+      iadst8_sse2(in);
+      iadst8_sse2(in);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 5);
+  in[1] = _mm_srai_epi16(in[1], 5);
+  in[2] = _mm_srai_epi16(in[2], 5);
+  in[3] = _mm_srai_epi16(in[3], 5);
+  in[4] = _mm_srai_epi16(in[4], 5);
+  in[5] = _mm_srai_epi16(in[5], 5);
+  in[6] = _mm_srai_epi16(in[6], 5);
+  in[7] = _mm_srai_epi16(in[7], 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+  RECON_AND_STORE(dest + 4 * stride, in[4]);
+  RECON_AND_STORE(dest + 5 * stride, in[5]);
+  RECON_AND_STORE(dest + 6 * stride, in[6]);
+  RECON_AND_STORE(dest + 7 * stride, in[7]);
+}
+
+void vp9_idct8x8_12_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // Rows. Load 4-row input data.
+  in0 = _mm_load_si128((const __m128i *)input);
+  in1 = _mm_load_si128((const __m128i *)(input + 8 * 1));
+  in2 = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in3 = _mm_load_si128((const __m128i *)(input + 8 * 3));
+
+  // 8x4 Transpose
+  TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
+  // Stage1
+  {
+    const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
+    const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_17, stg1_0);
+    tmp2 = _mm_madd_epi16(lo_17, stg1_1);
+    tmp4 = _mm_madd_epi16(lo_35, stg1_2);
+    tmp6 = _mm_madd_epi16(lo_35, stg1_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+    stp1_5 = _mm_packs_epi32(tmp4, tmp6);
+  }
+
+  // Stage2
+  {
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
+    const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_04, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_04, stg2_1);
+    tmp4 = _mm_madd_epi16(lo_26, stg2_2);
+    tmp6 = _mm_madd_epi16(lo_26, stg2_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp2_0 = _mm_packs_epi32(tmp0, tmp2);
+    stp2_2 = _mm_packs_epi32(tmp6, tmp4);
+
+    tmp0 = _mm_adds_epi16(stp1_4, stp1_5);
+    tmp1 = _mm_subs_epi16(stp1_4, stp1_5);
+
+    stp2_4 = tmp0;
+    stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
+
+    tmp4 = _mm_adds_epi16(stp2_0, stp2_2);
+    tmp6 = _mm_subs_epi16(stp2_0, stp2_2);
+
+    stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
+    stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
+
+    tmp0 = _mm_madd_epi16(lo_56, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0);  // stg3_1 = stg2_0
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_5 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  tmp0 = _mm_adds_epi16(stp1_3, stp2_4);
+  tmp1 = _mm_adds_epi16(stp1_2, stp1_5);
+  tmp2 = _mm_subs_epi16(stp1_3, stp2_4);
+  tmp3 = _mm_subs_epi16(stp1_2, stp1_5);
+
+  TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
+
+  IDCT8(in0, in1, in2, in3, zero, zero, zero, zero,
+        in0, in1, in2, in3, in4, in5, in6, in7);
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+#define IDCT16 \
+  /* Stage2 */ \
+  { \
+    const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]); \
+    const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]); \
+    const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]);   \
+    const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]);   \
+    const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]); \
+    const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]); \
+    const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]); \
+    const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, \
+                           stg2_0, stg2_1, stg2_2, stg2_3, \
+                           stp2_8, stp2_15, stp2_9, stp2_14) \
+    \
+    MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, \
+                           stg2_4, stg2_5, stg2_6, stg2_7, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  } \
+    \
+  /* Stage3 */ \
+  { \
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]); \
+    const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]); \
+    const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]); \
+    const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]); \
+    \
+    MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, \
+                           stg3_0, stg3_1, stg3_2, stg3_3, \
+                           stp1_4, stp1_7, stp1_5, stp1_6) \
+    \
+    stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9);  \
+    stp1_9 = _mm_sub_epi16(stp2_8, stp2_9);    \
+    stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+    stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+    \
+    stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13); \
+    stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+    stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  } \
+  \
+  /* Stage4 */ \
+  { \
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]); \
+    const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]); \
+    const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]); \
+    const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]); \
+    \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    \
+    MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, \
+                           stg4_0, stg4_1, stg4_2, stg4_3, \
+                           stp2_0, stp2_1, stp2_2, stp2_3) \
+    \
+    stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+    stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+    stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+    \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                           stg4_4, stg4_5, stg4_6, stg4_7, \
+                           stp2_9, stp2_14, stp2_10, stp2_13) \
+  } \
+    \
+  /* Stage5 */ \
+  { \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+    \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+    stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+    stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+    \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+    \
+    tmp0 = _mm_add_epi32(tmp0, rounding); \
+    tmp1 = _mm_add_epi32(tmp1, rounding); \
+    tmp2 = _mm_add_epi32(tmp2, rounding); \
+    tmp3 = _mm_add_epi32(tmp3, rounding); \
+    \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+    \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+    \
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+    \
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+  } \
+    \
+  /* Stage6 */ \
+  { \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+    \
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+    \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                           stg6_0, stg4_0, stg6_0, stg4_0, \
+                           stp2_10, stp2_13, stp2_11, stp2_12) \
+  }
+
+#define IDCT16_10 \
+    /* Stage2 */ \
+    { \
+      const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero); \
+      const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero); \
+      const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]); \
+      const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]); \
+      \
+      MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, \
+                             stg2_0, stg2_1, stg2_6, stg2_7, \
+                             stp1_8_0, stp1_15, stp1_11, stp1_12_0) \
+    } \
+      \
+    /* Stage3 */ \
+    { \
+      const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero); \
+      const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, \
+                               stg3_0, stg3_1,  \
+                               stp2_4, stp2_7) \
+      \
+      stp1_9  =  stp1_8_0; \
+      stp1_10 =  stp1_11;  \
+      \
+      stp1_13 = stp1_12_0; \
+      stp1_14 = stp1_15;   \
+    } \
+    \
+    /* Stage4 */ \
+    { \
+      const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero); \
+      const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero); \
+      \
+      const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+      const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      \
+      MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, \
+                               stg4_0, stg4_1, \
+                               stp1_0, stp1_1) \
+      stp2_5 = stp2_4; \
+      stp2_6 = stp2_7; \
+      \
+      MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, \
+                             stg4_4, stg4_5, stg4_6, stg4_7, \
+                             stp2_9, stp2_14, stp2_10, stp2_13) \
+    } \
+      \
+    /* Stage5 */ \
+    { \
+      const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+      const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+      \
+      stp1_2 = stp1_1; \
+      stp1_3 = stp1_0; \
+      \
+      tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+      tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+      tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+      tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+      \
+      tmp0 = _mm_add_epi32(tmp0, rounding); \
+      tmp1 = _mm_add_epi32(tmp1, rounding); \
+      tmp2 = _mm_add_epi32(tmp2, rounding); \
+      tmp3 = _mm_add_epi32(tmp3, rounding); \
+      \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+      \
+      stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+      stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+      \
+      stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);  \
+      stp1_9 = _mm_add_epi16(stp2_9, stp2_10);    \
+      stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);   \
+      stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11); \
+      \
+      stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0); \
+      stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);   \
+      stp1_14 = _mm_add_epi16(stp2_14, stp2_13);   \
+      stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0); \
+    } \
+      \
+    /* Stage6 */ \
+    { \
+      const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+      const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+      const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+      const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+      \
+      stp2_0 = _mm_add_epi16(stp1_0, stp2_7); \
+      stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+      stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+      stp2_3 = _mm_add_epi16(stp1_3, stp2_4); \
+      stp2_4 = _mm_sub_epi16(stp1_3, stp2_4); \
+      stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+      stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+      stp2_7 = _mm_sub_epi16(stp1_0, stp2_7); \
+      \
+      MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                             stg6_0, stg4_0, stg6_0, stg4_0, \
+                             stp2_10, stp2_13, stp2_11, stp2_12) \
+    }
+
+void vp9_idct16x16_256_add_sse2(const int16_t *input, uint8_t *dest,
+                                int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[16], l[16], r[16], *curr1;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  curr1 = l;
+  for (i = 0; i < 2; i++) {
+    // 1-D idct
+
+    // Load input data.
+    in[0] = _mm_load_si128((const __m128i *)input);
+    in[8] = _mm_load_si128((const __m128i *)(input + 8 * 1));
+    in[1] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+    in[9] = _mm_load_si128((const __m128i *)(input + 8 * 3));
+    in[2] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+    in[10] = _mm_load_si128((const __m128i *)(input + 8 * 5));
+    in[3] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+    in[11] = _mm_load_si128((const __m128i *)(input + 8 * 7));
+    in[4] = _mm_load_si128((const __m128i *)(input + 8 * 8));
+    in[12] = _mm_load_si128((const __m128i *)(input + 8 * 9));
+    in[5] = _mm_load_si128((const __m128i *)(input + 8 * 10));
+    in[13] = _mm_load_si128((const __m128i *)(input + 8 * 11));
+    in[6] = _mm_load_si128((const __m128i *)(input + 8 * 12));
+    in[14] = _mm_load_si128((const __m128i *)(input + 8 * 13));
+    in[7] = _mm_load_si128((const __m128i *)(input + 8 * 14));
+    in[15] = _mm_load_si128((const __m128i *)(input + 8 * 15));
+
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+
+    IDCT16
+
+    // Stage7
+    curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
+    curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
+    curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
+    curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
+    curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
+    curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
+    curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
+    curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
+    curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    curr1 = r;
+    input += 128;
+  }
+  for (i = 0; i < 2; i++) {
+    int j;
+    // 1-D idct
+    array_transpose_8x8(l + i * 8, in);
+    array_transpose_8x8(r + i * 8, in + 8);
+
+    IDCT16
+
+    // 2-D
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp9_idct16x16_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 2; ++i) {
+    RECON_AND_STORE(dest +  0 * stride, dc_value);
+    RECON_AND_STORE(dest +  1 * stride, dc_value);
+    RECON_AND_STORE(dest +  2 * stride, dc_value);
+    RECON_AND_STORE(dest +  3 * stride, dc_value);
+    RECON_AND_STORE(dest +  4 * stride, dc_value);
+    RECON_AND_STORE(dest +  5 * stride, dc_value);
+    RECON_AND_STORE(dest +  6 * stride, dc_value);
+    RECON_AND_STORE(dest +  7 * stride, dc_value);
+    RECON_AND_STORE(dest +  8 * stride, dc_value);
+    RECON_AND_STORE(dest +  9 * stride, dc_value);
+    RECON_AND_STORE(dest + 10 * stride, dc_value);
+    RECON_AND_STORE(dest + 11 * stride, dc_value);
+    RECON_AND_STORE(dest + 12 * stride, dc_value);
+    RECON_AND_STORE(dest + 13 * stride, dc_value);
+    RECON_AND_STORE(dest + 14 * stride, dc_value);
+    RECON_AND_STORE(dest + 15 * stride, dc_value);
+    dest += 8;
+  }
+}
+
+static void iadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void idct16_8col(__m128i *in) {
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i v[16], u[16], s[16], t[16];
+
+  // stage 1
+  s[0] = in[0];
+  s[1] = in[8];
+  s[2] = in[4];
+  s[3] = in[12];
+  s[4] = in[2];
+  s[5] = in[10];
+  s[6] = in[6];
+  s[7] = in[14];
+  s[8] = in[1];
+  s[9] = in[9];
+  s[10] = in[5];
+  s[11] = in[13];
+  s[12] = in[3];
+  s[13] = in[11];
+  s[14] = in[7];
+  s[15] = in[15];
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[15]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[15]);
+  u[2] = _mm_unpacklo_epi16(s[9], s[14]);
+  u[3] = _mm_unpackhi_epi16(s[9], s[14]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[11], s[12]);
+  u[7] = _mm_unpackhi_epi16(s[11], s[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[8]  = _mm_packs_epi32(u[0], u[1]);
+  s[15] = _mm_packs_epi32(u[2], u[3]);
+  s[9]  = _mm_packs_epi32(u[4], u[5]);
+  s[14] = _mm_packs_epi32(u[6], u[7]);
+  s[10] = _mm_packs_epi32(u[8], u[9]);
+  s[13] = _mm_packs_epi32(u[10], u[11]);
+  s[11] = _mm_packs_epi32(u[12], u[13]);
+  s[12] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  t[0] = s[0];
+  t[1] = s[1];
+  t[2] = s[2];
+  t[3] = s[3];
+  u[0] = _mm_unpacklo_epi16(s[4], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[5], s[6]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[4] = _mm_packs_epi32(u[0], u[1]);
+  t[7] = _mm_packs_epi32(u[2], u[3]);
+  t[5] = _mm_packs_epi32(u[4], u[5]);
+  t[6] = _mm_packs_epi32(u[6], u[7]);
+  t[8] = _mm_add_epi16(s[8], s[9]);
+  t[9] = _mm_sub_epi16(s[8], s[9]);
+  t[10] = _mm_sub_epi16(s[11], s[10]);
+  t[11] = _mm_add_epi16(s[10], s[11]);
+  t[12] = _mm_add_epi16(s[12], s[13]);
+  t[13] = _mm_sub_epi16(s[12], s[13]);
+  t[14] = _mm_sub_epi16(s[15], s[14]);
+  t[15] = _mm_add_epi16(s[14], s[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(t[0], t[1]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[1]);
+  u[2] = _mm_unpacklo_epi16(t[2], t[3]);
+  u[3] = _mm_unpackhi_epi16(t[2], t[3]);
+  u[4] = _mm_unpacklo_epi16(t[9], t[14]);
+  u[5] = _mm_unpackhi_epi16(t[9], t[14]);
+  u[6] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[7] = _mm_unpackhi_epi16(t[10], t[13]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_add_epi16(t[4], t[5]);
+  s[5] = _mm_sub_epi16(t[4], t[5]);
+  s[6] = _mm_sub_epi16(t[7], t[6]);
+  s[7] = _mm_add_epi16(t[6], t[7]);
+  s[8] = t[8];
+  s[15] = t[15];
+  s[9]  = _mm_packs_epi32(u[8], u[9]);
+  s[14] = _mm_packs_epi32(u[10], u[11]);
+  s[10] = _mm_packs_epi32(u[12], u[13]);
+  s[13] = _mm_packs_epi32(u[14], u[15]);
+  s[11] = t[11];
+  s[12] = t[12];
+
+  // stage 5
+  t[0] = _mm_add_epi16(s[0], s[3]);
+  t[1] = _mm_add_epi16(s[1], s[2]);
+  t[2] = _mm_sub_epi16(s[1], s[2]);
+  t[3] = _mm_sub_epi16(s[0], s[3]);
+  t[4] = s[4];
+  t[7] = s[7];
+
+  u[0] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[1] = _mm_unpackhi_epi16(s[5], s[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  t[5] = _mm_packs_epi32(u[0], u[1]);
+  t[6] = _mm_packs_epi32(u[2], u[3]);
+
+  t[8] = _mm_add_epi16(s[8], s[11]);
+  t[9] = _mm_add_epi16(s[9], s[10]);
+  t[10] = _mm_sub_epi16(s[9], s[10]);
+  t[11] = _mm_sub_epi16(s[8], s[11]);
+  t[12] = _mm_sub_epi16(s[15], s[12]);
+  t[13] = _mm_sub_epi16(s[14], s[13]);
+  t[14] = _mm_add_epi16(s[13], s[14]);
+  t[15] = _mm_add_epi16(s[12], s[15]);
+
+  // stage 6
+  s[0] = _mm_add_epi16(t[0], t[7]);
+  s[1] = _mm_add_epi16(t[1], t[6]);
+  s[2] = _mm_add_epi16(t[2], t[5]);
+  s[3] = _mm_add_epi16(t[3], t[4]);
+  s[4] = _mm_sub_epi16(t[3], t[4]);
+  s[5] = _mm_sub_epi16(t[2], t[5]);
+  s[6] = _mm_sub_epi16(t[1], t[6]);
+  s[7] = _mm_sub_epi16(t[0], t[7]);
+  s[8] = t[8];
+  s[9] = t[9];
+
+  u[0] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[1] = _mm_unpackhi_epi16(t[10], t[13]);
+  u[2] = _mm_unpacklo_epi16(t[11], t[12]);
+  u[3] = _mm_unpackhi_epi16(t[11], t[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  s[10] = _mm_packs_epi32(u[0], u[1]);
+  s[13] = _mm_packs_epi32(u[2], u[3]);
+  s[11] = _mm_packs_epi32(u[4], u[5]);
+  s[12] = _mm_packs_epi32(u[6], u[7]);
+  s[14] = t[14];
+  s[15] = t[15];
+
+  // stage 7
+  in[0] = _mm_add_epi16(s[0], s[15]);
+  in[1] = _mm_add_epi16(s[1], s[14]);
+  in[2] = _mm_add_epi16(s[2], s[13]);
+  in[3] = _mm_add_epi16(s[3], s[12]);
+  in[4] = _mm_add_epi16(s[4], s[11]);
+  in[5] = _mm_add_epi16(s[5], s[10]);
+  in[6] = _mm_add_epi16(s[6], s[9]);
+  in[7] = _mm_add_epi16(s[7], s[8]);
+  in[8] = _mm_sub_epi16(s[7], s[8]);
+  in[9] = _mm_sub_epi16(s[6], s[9]);
+  in[10] = _mm_sub_epi16(s[5], s[10]);
+  in[11] = _mm_sub_epi16(s[4], s[11]);
+  in[12] = _mm_sub_epi16(s[3], s[12]);
+  in[13] = _mm_sub_epi16(s[2], s[13]);
+  in[14] = _mm_sub_epi16(s[1], s[14]);
+  in[15] = _mm_sub_epi16(s[0], s[15]);
+}
+
+static void idct16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  idct16_8col(in0);
+  idct16_8col(in1);
+}
+
+static void iadst16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  iadst16_8col(in0);
+  iadst16_8col(in1);
+}
+
+void vp9_iht16x16_256_add_sse2(const int16_t *input, uint8_t *dest, int stride,
+                               int tx_type) {
+  __m128i in0[16], in1[16];
+
+  load_buffer_8x16(input, in0);
+  input += 8;
+  load_buffer_8x16(input, in1);
+
+  switch (tx_type) {
+    case 0:  // DCT_DCT
+      idct16_sse2(in0, in1);
+      idct16_sse2(in0, in1);
+      break;
+    case 1:  // ADST_DCT
+      idct16_sse2(in0, in1);
+      iadst16_sse2(in0, in1);
+      break;
+    case 2:  // DCT_ADST
+      iadst16_sse2(in0, in1);
+      idct16_sse2(in0, in1);
+      break;
+    case 3:  // ADST_ADST
+      iadst16_sse2(in0, in1);
+      iadst16_sse2(in0, in1);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+
+  write_buffer_8x16(dest, in0, stride);
+  dest += 8;
+  write_buffer_8x16(dest, in1, stride);
+}
+
+void vp9_idct16x16_10_add_sse2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i in[16], l[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+  // First 1-D inverse DCT
+  // Load input data.
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 8 * 2));
+  in[2] = _mm_load_si128((const __m128i *)(input + 8 * 4));
+  in[3] = _mm_load_si128((const __m128i *)(input + 8 * 6));
+
+  TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
+
+  // Stage2
+  {
+    const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
+    const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
+
+    tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
+    tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
+    tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp2_8  = _mm_packs_epi32(tmp0, tmp2);
+    stp2_11 = _mm_packs_epi32(tmp5, tmp7);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
+
+    tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
+    stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  {
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
+
+    tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
+    tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
+    tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp1_0 = _mm_packs_epi32(tmp0, tmp0);
+    stp1_1 = _mm_packs_epi32(tmp2, tmp2);
+    stp2_9 = _mm_packs_epi32(tmp1, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp5, tmp7);
+
+    stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
+  }
+
+  // Stage5 and Stage6
+  {
+    tmp0 = _mm_add_epi16(stp2_8, stp2_11);
+    tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
+    tmp2 = _mm_add_epi16(stp2_9, stp2_10);
+    tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
+
+    stp1_9  = _mm_unpacklo_epi64(tmp2, zero);
+    stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
+    stp1_8  = _mm_unpacklo_epi64(tmp0, zero);
+    stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
+
+    stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
+    stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
+    stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
+    stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
+  }
+
+  // Stage6
+  {
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
+
+    tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
+    tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
+    tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
+    tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
+    tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
+    tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
+
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_6 = _mm_packs_epi32(tmp3, tmp1);
+
+    stp2_10 = _mm_packs_epi32(tmp0, zero);
+    stp2_13 = _mm_packs_epi32(tmp2, zero);
+    stp2_11 = _mm_packs_epi32(tmp4, zero);
+    stp2_12 = _mm_packs_epi32(tmp6, zero);
+
+    tmp0 = _mm_add_epi16(stp1_0, stp1_4);
+    tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
+    tmp2 = _mm_add_epi16(stp1_1, stp1_6);
+    tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
+
+    stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
+    stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
+    stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
+    stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
+    stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage7. Left 8x16 only.
+  l[0] = _mm_add_epi16(stp2_0, stp1_15);
+  l[1] = _mm_add_epi16(stp2_1, stp1_14);
+  l[2] = _mm_add_epi16(stp2_2, stp2_13);
+  l[3] = _mm_add_epi16(stp2_3, stp2_12);
+  l[4] = _mm_add_epi16(stp2_4, stp2_11);
+  l[5] = _mm_add_epi16(stp2_5, stp2_10);
+  l[6] = _mm_add_epi16(stp2_6, stp1_9);
+  l[7] = _mm_add_epi16(stp2_7, stp1_8);
+  l[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  l[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  l[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  l[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  l[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  l[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  l[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  l[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+  // Second 1-D inverse transform, performed per 8x16 block
+  for (i = 0; i < 2; i++) {
+    int j;
+    array_transpose_4X8(l + 8 * i, in);
+
+    IDCT16_10
+
+    // Stage7
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+#define LOAD_DQCOEFF(reg, input) \
+  {  \
+    reg = _mm_load_si128((const __m128i *) input); \
+    input += 8; \
+  }  \
+
+#define IDCT32_34 \
+/* Stage1 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero); \
+  \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(zero, in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero); \
+  \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, \
+                         stg1_1, stp1_16, stp1_31); \
+  MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, \
+                         stg1_7, stp1_19, stp1_28); \
+  MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, \
+                         stg1_9, stp1_20, stp1_27); \
+  MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, \
+                         stg1_15, stp1_23, stp1_24); \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero); \
+  \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, \
+                         stg2_1, stp2_8, stp2_15); \
+  MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, \
+                         stg2_7, stp2_11, stp2_12); \
+  \
+  stp2_16 = stp1_16; \
+  stp2_19 = stp1_19; \
+  \
+  stp2_20 = stp1_20; \
+  stp2_23 = stp1_23; \
+  \
+  stp2_24 = stp1_24; \
+  stp2_27 = stp1_27; \
+  \
+  stp2_28 = stp1_28; \
+  stp2_31 = stp1_31; \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, \
+                         stg3_1, stp1_4, stp1_7); \
+  \
+  stp1_8 = stp2_8; \
+  stp1_11 = stp2_11; \
+  stp1_12 = stp2_12; \
+  stp1_15 = stp2_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i zero = _mm_setzero_si128();\
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12); \
+  \
+  MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, \
+                         stg4_1, stp2_0, stp2_1); \
+  \
+  stp2_4 = stp1_4; \
+  stp2_5 = stp1_4; \
+  stp2_6 = stp1_7; \
+  stp2_7 = stp1_7; \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = stp2_0; \
+  stp1_1 = stp2_1; \
+  stp1_2 = stp2_1; \
+  stp1_3 = stp2_0; \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+
+#define IDCT32 \
+/* Stage1 */ \
+{ \
+  const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], in[31]); \
+  const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], in[31]); \
+  const __m128i lo_17_15 = _mm_unpacklo_epi16(in[17], in[15]); \
+  const __m128i hi_17_15 = _mm_unpackhi_epi16(in[17], in[15]); \
+  \
+  const __m128i lo_9_23 = _mm_unpacklo_epi16(in[9], in[23]); \
+  const __m128i hi_9_23 = _mm_unpackhi_epi16(in[9], in[23]); \
+  const __m128i lo_25_7= _mm_unpacklo_epi16(in[25], in[7]); \
+  const __m128i hi_25_7 = _mm_unpackhi_epi16(in[25], in[7]); \
+  \
+  const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], in[27]); \
+  const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], in[27]); \
+  const __m128i lo_21_11 = _mm_unpacklo_epi16(in[21], in[11]); \
+  const __m128i hi_21_11 = _mm_unpackhi_epi16(in[21], in[11]); \
+  \
+  const __m128i lo_13_19 = _mm_unpacklo_epi16(in[13], in[19]); \
+  const __m128i hi_13_19 = _mm_unpackhi_epi16(in[13], in[19]); \
+  const __m128i lo_29_3 = _mm_unpacklo_epi16(in[29], in[3]); \
+  const __m128i hi_29_3 = _mm_unpackhi_epi16(in[29], in[3]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0, \
+                         stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, \
+                         stp1_17, stp1_30) \
+  MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, \
+                         stg1_5, stg1_6, stg1_7, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8, \
+                         stg1_9, stg1_10, stg1_11, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12, \
+                         stg1_13, stg1_14, stg1_15, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+} \
+\
+/* Stage2 */ \
+{ \
+  const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], in[30]); \
+  const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], in[30]); \
+  const __m128i lo_18_14 = _mm_unpacklo_epi16(in[18], in[14]); \
+  const __m128i hi_18_14 = _mm_unpackhi_epi16(in[18], in[14]); \
+  \
+  const __m128i lo_10_22 = _mm_unpacklo_epi16(in[10], in[22]); \
+  const __m128i hi_10_22 = _mm_unpackhi_epi16(in[10], in[22]); \
+  const __m128i lo_26_6 = _mm_unpacklo_epi16(in[26], in[6]); \
+  const __m128i hi_26_6 = _mm_unpackhi_epi16(in[26], in[6]); \
+  \
+  MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0, \
+                         stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, \
+                         stp2_14) \
+  MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4, \
+                         stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, \
+                         stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_17); \
+  stp2_17 = _mm_sub_epi16(stp1_16, stp1_17); \
+  stp2_18 = _mm_sub_epi16(stp1_19, stp1_18); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_18); \
+  \
+  stp2_20 = _mm_add_epi16(stp1_20, stp1_21); \
+  stp2_21 = _mm_sub_epi16(stp1_20, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_23, stp1_22); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_22); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_25); \
+  stp2_25 = _mm_sub_epi16(stp1_24, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_27, stp1_26); \
+  stp2_27 = _mm_add_epi16(stp1_27, stp1_26); \
+  \
+  stp2_28 = _mm_add_epi16(stp1_28, stp1_29); \
+  stp2_29 = _mm_sub_epi16(stp1_28, stp1_29); \
+  stp2_30 = _mm_sub_epi16(stp1_31, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_31, stp1_30); \
+} \
+\
+/* Stage3 */ \
+{ \
+  const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], in[28]); \
+  const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], in[28]); \
+  const __m128i lo_20_12 = _mm_unpacklo_epi16(in[20], in[12]); \
+  const __m128i hi_20_12 = _mm_unpackhi_epi16(in[20], in[12]); \
+  \
+  const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30); \
+  const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  \
+  MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0, \
+                         stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, \
+                         stp1_6) \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_9); \
+  stp1_9 = _mm_sub_epi16(stp2_8, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_11, stp2_10); \
+  stp1_11 = _mm_add_epi16(stp2_11, stp2_10); \
+  stp1_12 = _mm_add_epi16(stp2_12, stp2_13); \
+  stp1_13 = _mm_sub_epi16(stp2_12, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_15, stp2_14); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_14); \
+  \
+  MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4, \
+                         stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, \
+                         stp1_18, stp1_29) \
+  MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8, \
+                         stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, \
+                         stp1_22, stp1_25) \
+  \
+  stp1_16 = stp2_16; \
+  stp1_31 = stp2_31; \
+  stp1_19 = stp2_19; \
+  stp1_20 = stp2_20; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_27 = stp2_27; \
+  stp1_28 = stp2_28; \
+} \
+\
+/* Stage4 */ \
+{ \
+  const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], in[16]); \
+  const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], in[16]); \
+  const __m128i lo_8_24 = _mm_unpacklo_epi16(in[8], in[24]); \
+  const __m128i hi_8_24 = _mm_unpackhi_epi16(in[8], in[24]); \
+  \
+  const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14); \
+  const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14); \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  \
+  MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, \
+                         stg4_1, stg4_2, stg4_3, stp2_0, stp2_1, \
+                         stp2_2, stp2_3) \
+  \
+  stp2_4 = _mm_add_epi16(stp1_4, stp1_5); \
+  stp2_5 = _mm_sub_epi16(stp1_4, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_7, stp1_6); \
+  stp2_7 = _mm_add_epi16(stp1_7, stp1_6); \
+  \
+  MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4, \
+                         stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, \
+                         stp2_10, stp2_13) \
+  \
+  stp2_8 = stp1_8; \
+  stp2_15 = stp1_15; \
+  stp2_11 = stp1_11; \
+  stp2_12 = stp1_12; \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_19); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_18); \
+  stp2_18 = _mm_sub_epi16(stp1_17, stp1_18); \
+  stp2_19 = _mm_sub_epi16(stp1_16, stp1_19); \
+  stp2_20 = _mm_sub_epi16(stp1_23, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_22, stp1_21); \
+  stp2_22 = _mm_add_epi16(stp1_22, stp1_21); \
+  stp2_23 = _mm_add_epi16(stp1_23, stp1_20); \
+  \
+  stp2_24 = _mm_add_epi16(stp1_24, stp1_27); \
+  stp2_25 = _mm_add_epi16(stp1_25, stp1_26); \
+  stp2_26 = _mm_sub_epi16(stp1_25, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_24, stp1_27); \
+  stp2_28 = _mm_sub_epi16(stp1_31, stp1_28); \
+  stp2_29 = _mm_sub_epi16(stp1_30, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_29, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_28, stp1_31); \
+} \
+\
+/* Stage5 */ \
+{ \
+  const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5); \
+  const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5); \
+  const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29); \
+  const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29); \
+  \
+  const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28); \
+  const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28); \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_3); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_2); \
+  stp1_2 = _mm_sub_epi16(stp2_1, stp2_2); \
+  stp1_3 = _mm_sub_epi16(stp2_0, stp2_3); \
+  \
+  tmp0 = _mm_madd_epi16(lo_6_5, stg4_1); \
+  tmp1 = _mm_madd_epi16(hi_6_5, stg4_1); \
+  tmp2 = _mm_madd_epi16(lo_6_5, stg4_0); \
+  tmp3 = _mm_madd_epi16(hi_6_5, stg4_0); \
+  \
+  tmp0 = _mm_add_epi32(tmp0, rounding); \
+  tmp1 = _mm_add_epi32(tmp1, rounding); \
+  tmp2 = _mm_add_epi32(tmp2, rounding); \
+  tmp3 = _mm_add_epi32(tmp3, rounding); \
+  \
+  tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  \
+  stp1_5 = _mm_packs_epi32(tmp0, tmp1); \
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3); \
+  \
+  stp1_4 = stp2_4; \
+  stp1_7 = stp2_7; \
+  \
+  stp1_8 = _mm_add_epi16(stp2_8, stp2_11); \
+  stp1_9 = _mm_add_epi16(stp2_9, stp2_10); \
+  stp1_10 = _mm_sub_epi16(stp2_9, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_8, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_15, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_14, stp2_13); \
+  stp1_14 = _mm_add_epi16(stp2_14, stp2_13); \
+  stp1_15 = _mm_add_epi16(stp2_15, stp2_12); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  \
+  MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4, \
+                         stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, \
+                         stp1_19, stp1_28) \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6, \
+                         stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  \
+  stp1_22 = stp2_22; \
+  stp1_23 = stp2_23; \
+  stp1_24 = stp2_24; \
+  stp1_25 = stp2_25; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+} \
+\
+/* Stage6 */ \
+{ \
+  const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13); \
+  const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13); \
+  const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12); \
+  const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12); \
+  \
+  stp2_0 = _mm_add_epi16(stp1_0, stp1_7); \
+  stp2_1 = _mm_add_epi16(stp1_1, stp1_6); \
+  stp2_2 = _mm_add_epi16(stp1_2, stp1_5); \
+  stp2_3 = _mm_add_epi16(stp1_3, stp1_4); \
+  stp2_4 = _mm_sub_epi16(stp1_3, stp1_4); \
+  stp2_5 = _mm_sub_epi16(stp1_2, stp1_5); \
+  stp2_6 = _mm_sub_epi16(stp1_1, stp1_6); \
+  stp2_7 = _mm_sub_epi16(stp1_0, stp1_7); \
+  \
+  stp2_8 = stp1_8; \
+  stp2_9 = stp1_9; \
+  stp2_14 = stp1_14; \
+  stp2_15 = stp1_15; \
+  \
+  MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, \
+                         stg6_0, stg4_0, stg6_0, stg4_0, stp2_10, \
+                         stp2_13, stp2_11, stp2_12) \
+  \
+  stp2_16 = _mm_add_epi16(stp1_16, stp1_23); \
+  stp2_17 = _mm_add_epi16(stp1_17, stp1_22); \
+  stp2_18 = _mm_add_epi16(stp1_18, stp1_21); \
+  stp2_19 = _mm_add_epi16(stp1_19, stp1_20); \
+  stp2_20 = _mm_sub_epi16(stp1_19, stp1_20); \
+  stp2_21 = _mm_sub_epi16(stp1_18, stp1_21); \
+  stp2_22 = _mm_sub_epi16(stp1_17, stp1_22); \
+  stp2_23 = _mm_sub_epi16(stp1_16, stp1_23); \
+  \
+  stp2_24 = _mm_sub_epi16(stp1_31, stp1_24); \
+  stp2_25 = _mm_sub_epi16(stp1_30, stp1_25); \
+  stp2_26 = _mm_sub_epi16(stp1_29, stp1_26); \
+  stp2_27 = _mm_sub_epi16(stp1_28, stp1_27); \
+  stp2_28 = _mm_add_epi16(stp1_27, stp1_28); \
+  stp2_29 = _mm_add_epi16(stp1_26, stp1_29); \
+  stp2_30 = _mm_add_epi16(stp1_25, stp1_30); \
+  stp2_31 = _mm_add_epi16(stp1_24, stp1_31); \
+} \
+\
+/* Stage7 */ \
+{ \
+  const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27); \
+  const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27); \
+  const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26); \
+  const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26); \
+  \
+  const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25); \
+  const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25); \
+  const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24); \
+  const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24); \
+  \
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_15); \
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_14); \
+  stp1_2 = _mm_add_epi16(stp2_2, stp2_13); \
+  stp1_3 = _mm_add_epi16(stp2_3, stp2_12); \
+  stp1_4 = _mm_add_epi16(stp2_4, stp2_11); \
+  stp1_5 = _mm_add_epi16(stp2_5, stp2_10); \
+  stp1_6 = _mm_add_epi16(stp2_6, stp2_9); \
+  stp1_7 = _mm_add_epi16(stp2_7, stp2_8); \
+  stp1_8 = _mm_sub_epi16(stp2_7, stp2_8); \
+  stp1_9 = _mm_sub_epi16(stp2_6, stp2_9); \
+  stp1_10 = _mm_sub_epi16(stp2_5, stp2_10); \
+  stp1_11 = _mm_sub_epi16(stp2_4, stp2_11); \
+  stp1_12 = _mm_sub_epi16(stp2_3, stp2_12); \
+  stp1_13 = _mm_sub_epi16(stp2_2, stp2_13); \
+  stp1_14 = _mm_sub_epi16(stp2_1, stp2_14); \
+  stp1_15 = _mm_sub_epi16(stp2_0, stp2_15); \
+  \
+  stp1_16 = stp2_16; \
+  stp1_17 = stp2_17; \
+  stp1_18 = stp2_18; \
+  stp1_19 = stp2_19; \
+  \
+  MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, \
+                         stp1_21, stp1_26) \
+  MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0, \
+                         stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, \
+                         stp1_23, stp1_24) \
+  \
+  stp1_28 = stp2_28; \
+  stp1_29 = stp2_29; \
+  stp1_30 = stp2_30; \
+  stp1_31 = stp2_31; \
+}
+
+// Only upper-left 8x8 has non-zero coeff
+void vp9_idct32x32_34_add_sse2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[32];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data. Only need to load the top left 8x8 block.
+  in[0] = _mm_load_si128((const __m128i *)input);
+  in[1] = _mm_load_si128((const __m128i *)(input + 32));
+  in[2] = _mm_load_si128((const __m128i *)(input + 64));
+  in[3] = _mm_load_si128((const __m128i *)(input + 96));
+  in[4] = _mm_load_si128((const __m128i *)(input + 128));
+  in[5] = _mm_load_si128((const __m128i *)(input + 160));
+  in[6] = _mm_load_si128((const __m128i *)(input + 192));
+  in[7] = _mm_load_si128((const __m128i *)(input + 224));
+
+  for (i = 8; i < 32; ++i) {
+    in[i] = _mm_setzero_si128();
+  }
+
+  array_transpose_8x8(in, in);
+  // TODO(hkuang): Following transposes are unnecessary. But remove them will
+  // lead to performance drop on some devices.
+  array_transpose_8x8(in + 8, in + 8);
+  array_transpose_8x8(in + 16, in + 16);
+  array_transpose_8x8(in + 24, in + 24);
+
+  IDCT32_34
+
+  // 1_D: Store 32 intermediate results for each 8x32 block.
+  col[0] = _mm_add_epi16(stp1_0, stp1_31);
+  col[1] = _mm_add_epi16(stp1_1, stp1_30);
+  col[2] = _mm_add_epi16(stp1_2, stp1_29);
+  col[3] = _mm_add_epi16(stp1_3, stp1_28);
+  col[4] = _mm_add_epi16(stp1_4, stp1_27);
+  col[5] = _mm_add_epi16(stp1_5, stp1_26);
+  col[6] = _mm_add_epi16(stp1_6, stp1_25);
+  col[7] = _mm_add_epi16(stp1_7, stp1_24);
+  col[8] = _mm_add_epi16(stp1_8, stp1_23);
+  col[9] = _mm_add_epi16(stp1_9, stp1_22);
+  col[10] = _mm_add_epi16(stp1_10, stp1_21);
+  col[11] = _mm_add_epi16(stp1_11, stp1_20);
+  col[12] = _mm_add_epi16(stp1_12, stp1_19);
+  col[13] = _mm_add_epi16(stp1_13, stp1_18);
+  col[14] = _mm_add_epi16(stp1_14, stp1_17);
+  col[15] = _mm_add_epi16(stp1_15, stp1_16);
+  col[16] = _mm_sub_epi16(stp1_15, stp1_16);
+  col[17] = _mm_sub_epi16(stp1_14, stp1_17);
+  col[18] = _mm_sub_epi16(stp1_13, stp1_18);
+  col[19] = _mm_sub_epi16(stp1_12, stp1_19);
+  col[20] = _mm_sub_epi16(stp1_11, stp1_20);
+  col[21] = _mm_sub_epi16(stp1_10, stp1_21);
+  col[22] = _mm_sub_epi16(stp1_9, stp1_22);
+  col[23] = _mm_sub_epi16(stp1_8, stp1_23);
+  col[24] = _mm_sub_epi16(stp1_7, stp1_24);
+  col[25] = _mm_sub_epi16(stp1_6, stp1_25);
+  col[26] = _mm_sub_epi16(stp1_5, stp1_26);
+  col[27] = _mm_sub_epi16(stp1_4, stp1_27);
+  col[28] = _mm_sub_epi16(stp1_3, stp1_28);
+  col[29] = _mm_sub_epi16(stp1_2, stp1_29);
+  col[30] = _mm_sub_epi16(stp1_1, stp1_30);
+  col[31] = _mm_sub_epi16(stp1_0, stp1_31);
+  for (i = 0; i < 4; i++) {
+    int j;
+    const __m128i zero = _mm_setzero_si128();
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + i * 8, in);
+    IDCT32_34
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp9_idct32x32_1024_add_sse2(const int16_t *input, uint8_t *dest,
+                                 int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[128], zero_idx[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+          stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+          stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22,
+          stp1_23, stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29,
+          stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+          stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+          stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22,
+          stp2_23, stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29,
+          stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i, j, i32;
+
+  for (i = 0; i < 4; i++) {
+    i32 = (i << 5);
+    // First 1-D idct
+    // Load input data.
+    LOAD_DQCOEFF(in[0], input);
+    LOAD_DQCOEFF(in[8], input);
+    LOAD_DQCOEFF(in[16], input);
+    LOAD_DQCOEFF(in[24], input);
+    LOAD_DQCOEFF(in[1], input);
+    LOAD_DQCOEFF(in[9], input);
+    LOAD_DQCOEFF(in[17], input);
+    LOAD_DQCOEFF(in[25], input);
+    LOAD_DQCOEFF(in[2], input);
+    LOAD_DQCOEFF(in[10], input);
+    LOAD_DQCOEFF(in[18], input);
+    LOAD_DQCOEFF(in[26], input);
+    LOAD_DQCOEFF(in[3], input);
+    LOAD_DQCOEFF(in[11], input);
+    LOAD_DQCOEFF(in[19], input);
+    LOAD_DQCOEFF(in[27], input);
+
+    LOAD_DQCOEFF(in[4], input);
+    LOAD_DQCOEFF(in[12], input);
+    LOAD_DQCOEFF(in[20], input);
+    LOAD_DQCOEFF(in[28], input);
+    LOAD_DQCOEFF(in[5], input);
+    LOAD_DQCOEFF(in[13], input);
+    LOAD_DQCOEFF(in[21], input);
+    LOAD_DQCOEFF(in[29], input);
+    LOAD_DQCOEFF(in[6], input);
+    LOAD_DQCOEFF(in[14], input);
+    LOAD_DQCOEFF(in[22], input);
+    LOAD_DQCOEFF(in[30], input);
+    LOAD_DQCOEFF(in[7], input);
+    LOAD_DQCOEFF(in[15], input);
+    LOAD_DQCOEFF(in[23], input);
+    LOAD_DQCOEFF(in[31], input);
+
+    // checking if all entries are zero
+    zero_idx[0] = _mm_or_si128(in[0], in[1]);
+    zero_idx[1] = _mm_or_si128(in[2], in[3]);
+    zero_idx[2] = _mm_or_si128(in[4], in[5]);
+    zero_idx[3] = _mm_or_si128(in[6], in[7]);
+    zero_idx[4] = _mm_or_si128(in[8], in[9]);
+    zero_idx[5] = _mm_or_si128(in[10], in[11]);
+    zero_idx[6] = _mm_or_si128(in[12], in[13]);
+    zero_idx[7] = _mm_or_si128(in[14], in[15]);
+    zero_idx[8] = _mm_or_si128(in[16], in[17]);
+    zero_idx[9] = _mm_or_si128(in[18], in[19]);
+    zero_idx[10] = _mm_or_si128(in[20], in[21]);
+    zero_idx[11] = _mm_or_si128(in[22], in[23]);
+    zero_idx[12] = _mm_or_si128(in[24], in[25]);
+    zero_idx[13] = _mm_or_si128(in[26], in[27]);
+    zero_idx[14] = _mm_or_si128(in[28], in[29]);
+    zero_idx[15] = _mm_or_si128(in[30], in[31]);
+
+    zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+    zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
+
+    zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+
+    if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
+      col[i32 + 0] = _mm_setzero_si128();
+      col[i32 + 1] = _mm_setzero_si128();
+      col[i32 + 2] = _mm_setzero_si128();
+      col[i32 + 3] = _mm_setzero_si128();
+      col[i32 + 4] = _mm_setzero_si128();
+      col[i32 + 5] = _mm_setzero_si128();
+      col[i32 + 6] = _mm_setzero_si128();
+      col[i32 + 7] = _mm_setzero_si128();
+      col[i32 + 8] = _mm_setzero_si128();
+      col[i32 + 9] = _mm_setzero_si128();
+      col[i32 + 10] = _mm_setzero_si128();
+      col[i32 + 11] = _mm_setzero_si128();
+      col[i32 + 12] = _mm_setzero_si128();
+      col[i32 + 13] = _mm_setzero_si128();
+      col[i32 + 14] = _mm_setzero_si128();
+      col[i32 + 15] = _mm_setzero_si128();
+      col[i32 + 16] = _mm_setzero_si128();
+      col[i32 + 17] = _mm_setzero_si128();
+      col[i32 + 18] = _mm_setzero_si128();
+      col[i32 + 19] = _mm_setzero_si128();
+      col[i32 + 20] = _mm_setzero_si128();
+      col[i32 + 21] = _mm_setzero_si128();
+      col[i32 + 22] = _mm_setzero_si128();
+      col[i32 + 23] = _mm_setzero_si128();
+      col[i32 + 24] = _mm_setzero_si128();
+      col[i32 + 25] = _mm_setzero_si128();
+      col[i32 + 26] = _mm_setzero_si128();
+      col[i32 + 27] = _mm_setzero_si128();
+      col[i32 + 28] = _mm_setzero_si128();
+      col[i32 + 29] = _mm_setzero_si128();
+      col[i32 + 30] = _mm_setzero_si128();
+      col[i32 + 31] = _mm_setzero_si128();
+      continue;
+    }
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+    array_transpose_8x8(in + 16, in + 16);
+    array_transpose_8x8(in + 24, in + 24);
+
+    IDCT32
+
+    // 1_D: Store 32 intermediate results for each 8x32 block.
+    col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
+    col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
+    col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
+    col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
+    col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
+    col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
+    col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
+    col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
+    col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
+    col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
+    col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
+    col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
+    col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
+    col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
+    col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
+    col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
+    col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
+    col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
+    col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
+    col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
+    col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
+    col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
+    col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
+    col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
+    col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
+    col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
+    col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
+    col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
+    col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
+    col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
+    col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
+    col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
+  }
+  for (i = 0; i < 4; i++) {
+    // Second 1-D idct
+    j = i << 3;
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + j, in);
+    array_transpose_8x8(col + j + 32, in + 8);
+    array_transpose_8x8(col + j + 64, in + 16);
+    array_transpose_8x8(col + j + 96, in + 24);
+
+    IDCT32
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void vp9_idct32x32_1_add_sse2(const int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 4; ++i) {
+    int j;
+    for (j = 0; j < 32; ++j) {
+      RECON_AND_STORE(dest + j * stride, dc_value);
+    }
+    dest += 8;
+  }
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
+  __m128i ubounded, retval;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one);
+  ubounded = _mm_cmpgt_epi16(value, max);
+  retval = _mm_andnot_si128(ubounded, value);
+  ubounded = _mm_and_si128(ubounded, max);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
+  return retval;
+}
+
+void vp9_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  __m128i inptr[4];
+  __m128i sign_bits[2];
+  __m128i temp_mm, min_input, max_input;
+  int test;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  int optimised_cols = 0;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i max = _mm_set1_epi16(12043);
+  const __m128i min = _mm_set1_epi16(-12043);
+  // Load input into __m128i
+  inptr[0] = _mm_loadu_si128((const __m128i *)input);
+  inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
+  inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
+  inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
+
+  // Pack to 16 bits
+  inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
+  inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
+
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp_mm = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp_mm);
+
+  if (!test) {
+    // Do the row transform
+    idct4_sse2(inptr);
+
+    // Check the min & max values
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp_mm = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp_mm);
+
+    if (test) {
+      transpose_4x4(inptr);
+      sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
+      sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
+      inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
+      inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
+      inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
+      inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
+      _mm_storeu_si128((__m128i *)outptr, inptr[0]);
+      _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
+      _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
+      _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp9_highbd_idct4(input, outptr, bd);
+      input += 4;
+      outptr += 4;
+    }
+  }
+
+  if (optimised_cols) {
+    idct4_sse2(inptr);
+
+    // Final round and shift
+    inptr[0] = _mm_add_epi16(inptr[0], eight);
+    inptr[1] = _mm_add_epi16(inptr[1], eight);
+
+    inptr[0] = _mm_srai_epi16(inptr[0], 4);
+    inptr[1] = _mm_srai_epi16(inptr[1], 4);
+
+    // Reconstruction and Store
+    {
+      __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
+      __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
+      d0 = _mm_unpacklo_epi64(
+          d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
+      d2 = _mm_unpacklo_epi64(
+          d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
+      d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
+      d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
+      // store input0
+      _mm_storel_epi64((__m128i *)dest, d0);
+      // store input1
+      d0 = _mm_srli_si128(d0, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride), d0);
+      // store input2
+      _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
+      // store input3
+      d2 = _mm_srli_si128(d2, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[4], temp_out[4];
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = out[j * 4 + i];
+      vp9_highbd_idct4(temp_in, temp_out, bd);
+      for (j = 0; j < 4; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+      }
+    }
+  }
+}
+
+void vp9_highbd_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 8; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_8x8(inptr, inptr);
+      for (i = 0; i < 8; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 8; ++i) {
+      vp9_highbd_idct8(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vp9_highbd_idct8(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vp9_highbd_idct8x8_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[8];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  const __m128i max = _mm_set1_epi16(6201);
+  const __m128i min = _mm_set1_epi16(-6201);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 8; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 8 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 8 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // only first 4 row has non-zero coefs
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct8_sse2(inptr);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 8; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_4X8(inptr, inptr);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i + 1)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (2 * i)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp9_highbd_idct8(input, outptr, bd);
+      input += 8;
+      outptr += 8;
+    }
+  }
+
+  if (optimised_cols) {
+    idct8_sse2(inptr);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[8];
+      for (i = 0; i < 8; i++) {
+        inptr[i] = _mm_add_epi16(inptr[i], sixteen);
+        d[i] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        inptr[i] = _mm_srai_epi16(inptr[i], 5);
+        d[i] = clamp_high_sse2(_mm_adds_epi16(d[i], inptr[i]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[i]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[8], temp_out[8];
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j * 8 + i];
+      vp9_highbd_idct8(temp_in, temp_out, bd);
+      for (j = 0; j < 8; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 5), bd);
+      }
+    }
+  }
+}
+
+void vp9_highbd_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                       int stride, int bd) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 32; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform
+    idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 32; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      array_transpose_16x16(inptr, inptr + 16);
+      for (i = 0; i < 16; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 16; ++i) {
+      vp9_highbd_idct16(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vp9_highbd_idct16(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+
+void vp9_highbd_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                      int stride, int bd) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j, test;
+  __m128i inptr[32];
+  __m128i min_input, max_input, temp1, temp2, sign_bits;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i rounding = _mm_set1_epi16(32);
+  const __m128i max = _mm_set1_epi16(3155);
+  const __m128i min = _mm_set1_epi16(-3155);
+  int optimised_cols = 0;
+
+  // Load input into __m128i & pack to 16 bits
+  for (i = 0; i < 16; i++) {
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 4));
+    inptr[i] = _mm_packs_epi32(temp1, temp2);
+    temp1 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 8));
+    temp2 = _mm_loadu_si128((const __m128i *)(input + 16 * i + 12));
+    inptr[i + 16] = _mm_packs_epi32(temp1, temp2);
+  }
+
+  // Find the min & max for the row transform
+  // Since all non-zero dct coefficients are in upper-left 4x4 area,
+  // we only need to consider first 4 rows here.
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  for (i = 2; i < 4; i++) {
+    max_input = _mm_max_epi16(max_input, inptr[i]);
+    min_input = _mm_min_epi16(min_input, inptr[i]);
+  }
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp1 = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp1);
+
+  if (!test) {
+    // Do the row transform (N.B. This transposes inptr)
+    idct16_sse2(inptr, inptr + 16);
+
+    // Find the min & max for the column transform
+    // N.B. Only first 4 cols contain non-zero coeffs
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    for (i = 2; i < 16; i++) {
+      max_input = _mm_max_epi16(max_input, inptr[i]);
+      min_input = _mm_min_epi16(min_input, inptr[i]);
+    }
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp1 = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp1);
+
+    if (test) {
+      // Use fact only first 4 rows contain non-zero coeffs
+      array_transpose_8x8(inptr, inptr);
+      array_transpose_8x8(inptr + 8, inptr + 16);
+      for (i = 0; i < 4; i++) {
+        sign_bits = _mm_cmplt_epi16(inptr[i], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 1)), temp2);
+        sign_bits = _mm_cmplt_epi16(inptr[i + 16], zero);
+        temp1 = _mm_unpacklo_epi16(inptr[i + 16], sign_bits);
+        temp2 = _mm_unpackhi_epi16(inptr[i + 16], sign_bits);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 2)), temp1);
+        _mm_storeu_si128((__m128i *)(outptr + 4 * (i * 4 + 3)), temp2);
+      }
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      vp9_highbd_idct16(input, outptr, bd);
+      input += 16;
+      outptr += 16;
+    }
+  }
+
+  if (optimised_cols) {
+    idct16_sse2(inptr, inptr + 16);
+
+    // Final round & shift and Reconstruction and Store
+    {
+      __m128i d[2];
+      for (i = 0; i < 16; i++) {
+        inptr[i   ] = _mm_add_epi16(inptr[i   ], rounding);
+        inptr[i+16] = _mm_add_epi16(inptr[i+16], rounding);
+        d[0] = _mm_loadu_si128((const __m128i *)(dest + stride*i));
+        d[1] = _mm_loadu_si128((const __m128i *)(dest + stride*i + 8));
+        inptr[i   ] = _mm_srai_epi16(inptr[i   ], 6);
+        inptr[i+16] = _mm_srai_epi16(inptr[i+16], 6);
+        d[0] = clamp_high_sse2(_mm_add_epi16(d[0], inptr[i   ]), bd);
+        d[1] = clamp_high_sse2(_mm_add_epi16(d[1], inptr[i+16]), bd);
+        // Store
+        _mm_storeu_si128((__m128i *)(dest + stride*i), d[0]);
+        _mm_storeu_si128((__m128i *)(dest + stride*i + 8), d[1]);
+      }
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[16], temp_out[16];
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j * 16 + i];
+      vp9_highbd_idct16(temp_in, temp_out, bd);
+      for (j = 0; j < 16; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 6), bd);
+      }
+    }
+  }
+}
+
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h b/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h
new file mode 100644
index 000000000..984363d40
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_idct_intrin_sse2.h
@@ -0,0 +1,174 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_idct.h"
+
+// perform 8x8 transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+}
+
+#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1) \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+                                                        \
+    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1);  /* i1 i0 */  \
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1);  /* i3 i2 */  \
+  }
+
+static INLINE void array_transpose_4X8(__m128i *in, __m128i * out) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+
+  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in) {
+  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * 16));
+  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * 16));
+  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * 16));
+  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * 16));
+  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * 16));
+  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * 16));
+  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * 16));
+  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * 16));
+
+  in[8]  = _mm_load_si128((const __m128i *)(input + 8 * 16));
+  in[9]  = _mm_load_si128((const __m128i *)(input + 9 * 16));
+  in[10]  = _mm_load_si128((const __m128i *)(input + 10 * 16));
+  in[11]  = _mm_load_si128((const __m128i *)(input + 11 * 16));
+  in[12]  = _mm_load_si128((const __m128i *)(input + 12 * 16));
+  in[13]  = _mm_load_si128((const __m128i *)(input + 13 * 16));
+  in[14]  = _mm_load_si128((const __m128i *)(input + 14 * 16));
+  in[15]  = _mm_load_si128((const __m128i *)(input + 15 * 16));
+}
+
+#define RECON_AND_STORE(dest, in_x) \
+  {                                                     \
+     __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+      d0 = _mm_unpacklo_epi8(d0, zero); \
+      d0 = _mm_add_epi16(in_x, d0); \
+      d0 = _mm_packus_epi16(d0, d0); \
+      _mm_storel_epi64((__m128i *)(dest), d0); \
+  }
+
+static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1<<5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+  in[8] = _mm_adds_epi16(in[8], final_rounding);
+  in[9] = _mm_adds_epi16(in[9], final_rounding);
+  in[10] = _mm_adds_epi16(in[10], final_rounding);
+  in[11] = _mm_adds_epi16(in[11], final_rounding);
+  in[12] = _mm_adds_epi16(in[12], final_rounding);
+  in[13] = _mm_adds_epi16(in[13], final_rounding);
+  in[14] = _mm_adds_epi16(in[14], final_rounding);
+  in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+  in[8] = _mm_srai_epi16(in[8], 6);
+  in[9] = _mm_srai_epi16(in[9], 6);
+  in[10] = _mm_srai_epi16(in[10], 6);
+  in[11] = _mm_srai_epi16(in[11], 6);
+  in[12] = _mm_srai_epi16(in[12], 6);
+  in[13] = _mm_srai_epi16(in[13], 6);
+  in[14] = _mm_srai_epi16(in[14], 6);
+  in[15] = _mm_srai_epi16(in[15], 6);
+
+  RECON_AND_STORE(dest +  0 * stride, in[0]);
+  RECON_AND_STORE(dest +  1 * stride, in[1]);
+  RECON_AND_STORE(dest +  2 * stride, in[2]);
+  RECON_AND_STORE(dest +  3 * stride, in[3]);
+  RECON_AND_STORE(dest +  4 * stride, in[4]);
+  RECON_AND_STORE(dest +  5 * stride, in[5]);
+  RECON_AND_STORE(dest +  6 * stride, in[6]);
+  RECON_AND_STORE(dest +  7 * stride, in[7]);
+  RECON_AND_STORE(dest +  8 * stride, in[8]);
+  RECON_AND_STORE(dest +  9 * stride, in[9]);
+  RECON_AND_STORE(dest + 10 * stride, in[10]);
+  RECON_AND_STORE(dest + 11 * stride, in[11]);
+  RECON_AND_STORE(dest + 12 * stride, in[12]);
+  RECON_AND_STORE(dest + 13 * stride, in[13]);
+  RECON_AND_STORE(dest + 14 * stride, in[14]);
+  RECON_AND_STORE(dest + 15 * stride, in[15]);
+}
diff --git a/media/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm b/media/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm
new file mode 100644
index 000000000..2c1060710
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_idct_ssse3_x86_64.asm
@@ -0,0 +1,300 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the inverse transformation. Part
+; of the functions are originally derived from the ffmpeg project.
+; Note that the current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+pw_16:      times 8 dw 16
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_m%2_%1:  dw -%2,  %1, -%2,  %1, -%2,  %1, -%2,  %1
+%endmacro
+
+TRANSFORM_COEFFS    6270, 15137
+TRANSFORM_COEFFS    3196, 16069
+TRANSFORM_COEFFS   13623,  9102
+
+%macro PAIR_PP_COEFFS 2
+dpw_%1_%2:   dw  %1,  %1,  %1,  %1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MP_COEFFS 2
+dpw_m%1_%2:  dw -%1, -%1, -%1, -%1,  %2,  %2,  %2,  %2
+%endmacro
+
+%macro PAIR_MM_COEFFS 2
+dpw_m%1_m%2: dw -%1, -%1, -%1, -%1, -%2, -%2, -%2, -%2
+%endmacro
+
+PAIR_PP_COEFFS     30274, 12540
+PAIR_PP_COEFFS      6392, 32138
+PAIR_MP_COEFFS     18204, 27246
+
+PAIR_PP_COEFFS     12540, 12540
+PAIR_PP_COEFFS     30274, 30274
+PAIR_PP_COEFFS      6392,  6392
+PAIR_PP_COEFFS     32138, 32138
+PAIR_MM_COEFFS     18204, 18204
+PAIR_PP_COEFFS     27246, 27246
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_m%4_%3], [pw_%3_%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_m%4_%3], [pw_%3_%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+%macro IDCT8_1D 0
+  SUM_SUB          0,    4,    9
+  BUTTERFLY_4X     2,    6,    6270, 15137,  m8,  9,  10
+  pmulhrsw        m0,  m12
+  pmulhrsw        m4,  m12
+  BUTTERFLY_4X     1,    7,    3196, 16069,  m8,  9,  10
+  BUTTERFLY_4X     5,    3,   13623,  9102,  m8,  9,  10
+
+  SUM_SUB          1,    5,    9
+  SUM_SUB          7,    3,    9
+  SUM_SUB          0,    6,    9
+  SUM_SUB          4,    2,    9
+  SUM_SUB          3,    5,    9
+  pmulhrsw        m3,  m12
+  pmulhrsw        m5,  m12
+
+  SUM_SUB          0,    7,    9
+  SUM_SUB          4,    3,    9
+  SUM_SUB          2,    5,    9
+  SUM_SUB          6,    1,    9
+
+  SWAP             3,    6
+  SWAP             1,    4
+%endmacro
+
+; This macro handles 8 pixels per line
+%macro ADD_STORE_8P_2X 5;  src1, src2, tmp1, tmp2, zero
+  paddw           m%1, m11
+  paddw           m%2, m11
+  psraw           m%1, 5
+  psraw           m%2, 5
+
+  movh            m%3, [outputq]
+  movh            m%4, [outputq + strideq]
+  punpcklbw       m%3, m%5
+  punpcklbw       m%4, m%5
+  paddw           m%3, m%1
+  paddw           m%4, m%2
+  packuswb        m%3, m%5
+  packuswb        m%4, m%5
+  movh               [outputq], m%3
+  movh     [outputq + strideq], m%4
+%endmacro
+
+INIT_XMM ssse3
+; full inverse 8x8 2D-DCT transform
+cglobal idct8x8_64_add, 3, 5, 13, input, output, stride
+  mova     m8, [pd_8192]
+  mova    m11, [pw_16]
+  mova    m12, [pw_11585x2]
+
+  lea      r3, [2 * strideq]
+
+  mova     m0, [inputq +   0]
+  mova     m1, [inputq +  16]
+  mova     m2, [inputq +  32]
+  mova     m3, [inputq +  48]
+  mova     m4, [inputq +  64]
+  mova     m5, [inputq +  80]
+  mova     m6, [inputq +  96]
+  mova     m7, [inputq + 112]
+
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+  TRANSPOSE8X8  0, 1, 2, 3, 4, 5, 6, 7, 9
+  IDCT8_1D
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+; inverse 8x8 2D-DCT transform with only first 10 coeffs non-zero
+cglobal idct8x8_12_add, 3, 5, 13, input, output, stride
+  mova       m8, [pd_8192]
+  mova      m11, [pw_16]
+  mova      m12, [pw_11585x2]
+
+  lea        r3, [2 * strideq]
+
+  mova       m0, [inputq +  0]
+  mova       m1, [inputq + 16]
+  mova       m2, [inputq + 32]
+  mova       m3, [inputq + 48]
+
+  punpcklwd  m0, m1
+  punpcklwd  m2, m3
+  punpckhdq  m9, m0, m2
+  punpckldq  m0, m2
+  SWAP       2, 9
+
+  ; m0 -> [0], [0]
+  ; m1 -> [1], [1]
+  ; m2 -> [2], [2]
+  ; m3 -> [3], [3]
+  punpckhqdq m10, m0, m0
+  punpcklqdq m0,  m0
+  punpckhqdq m9,  m2, m2
+  punpcklqdq m2,  m2
+  SWAP       1, 10
+  SWAP       3,  9
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m2, [dpw_30274_12540]
+  pmulhrsw   m1, [dpw_6392_32138]
+  pmulhrsw   m3, [dpw_m18204_27246]
+
+  SUM_SUB    0, 2, 9
+  SUM_SUB    1, 3, 9
+
+  punpcklqdq m9, m3, m3
+  punpckhqdq m5, m3, m9
+
+  SUM_SUB    3, 5, 9
+  punpckhqdq m5, m3
+  pmulhrsw   m5, m12
+
+  punpckhqdq m9, m1, m5
+  punpcklqdq m1, m5
+  SWAP       5, 9
+
+  SUM_SUB    0, 5, 9
+  SUM_SUB    2, 1, 9
+
+  punpckhqdq m3, m0, m0
+  punpckhqdq m4, m1, m1
+  punpckhqdq m6, m5, m5
+  punpckhqdq m7, m2, m2
+
+  punpcklwd  m0, m3
+  punpcklwd  m7, m2
+  punpcklwd  m1, m4
+  punpcklwd  m6, m5
+
+  punpckhdq  m4, m0, m7
+  punpckldq  m0, m7
+  punpckhdq  m10, m1, m6
+  punpckldq  m5, m1, m6
+
+  punpckhqdq m1, m0, m5
+  punpcklqdq m0, m5
+  punpckhqdq m3, m4, m10
+  punpcklqdq m2, m4, m10
+
+
+  pmulhrsw   m0, m12
+  pmulhrsw   m6, m2, [dpw_30274_30274]
+  pmulhrsw   m4, m2, [dpw_12540_12540]
+
+  pmulhrsw   m7, m1, [dpw_32138_32138]
+  pmulhrsw   m1, [dpw_6392_6392]
+  pmulhrsw   m5, m3, [dpw_m18204_m18204]
+  pmulhrsw   m3, [dpw_27246_27246]
+
+  mova       m2, m0
+  SUM_SUB    0, 6, 9
+  SUM_SUB    2, 4, 9
+  SUM_SUB    1, 5, 9
+  SUM_SUB    7, 3, 9
+
+  SUM_SUB    3, 5, 9
+  pmulhrsw   m3, m12
+  pmulhrsw   m5, m12
+
+  SUM_SUB    0, 7, 9
+  SUM_SUB    2, 3, 9
+  SUM_SUB    4, 5, 9
+  SUM_SUB    6, 1, 9
+
+  SWAP       3, 6
+  SWAP       1, 2
+  SWAP       2, 4
+
+
+  pxor    m12, m12
+  ADD_STORE_8P_2X  0, 1, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  2, 3, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  4, 5, 9, 10, 12
+  lea              outputq, [outputq + r3]
+  ADD_STORE_8P_2X  6, 7, 9, 10, 12
+
+  RET
+
+%endif
diff --git a/media/libvpx/vp9/common/x86/vp9_intrapred_sse2.asm b/media/libvpx/vp9/common/x86/vp9_intrapred_sse2.asm
new file mode 100644
index 000000000..22b573188
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_intrapred_sse2.asm
@@ -0,0 +1,667 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 8 dw 16
+pw_32: times 8 dw 32
+dc_128: times 16 db 128
+pw2_4:  times 8 dw 2
+pw2_8:  times 8 dw 4
+pw2_16:  times 8 dw 8
+pw2_32:  times 8 dw 16
+
+SECTION .text
+
+INIT_MMX sse
+cglobal dc_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [aboveq]
+  punpckldq             m0, [leftq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_left_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [leftq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_top_predictor_4x4, 4, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [aboveq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_8)]
+  psraw                 m0, 4
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_top_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_left_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  pshufw                m0, m0, 0x0
+  packuswb              m0, m0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_128_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movd     m0,        [GLOBAL(dc_128)]
+  movd    [dstq          ], m0
+  movd    [dstq+strideq  ], m0
+  movd    [dstq+strideq*2], m0
+  movd    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal dc_128_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    m0,        [GLOBAL(dc_128)]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_16)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+
+INIT_XMM sse2
+cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  pxor                  m2, m2
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  pxor                  m2, m2
+  mova                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+
+INIT_XMM sse2
+cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  psadbw                m3, m1
+  psadbw                m4, m1
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_32)]
+  psraw                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  mova                  m2, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+INIT_MMX sse
+cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movd                  m0, [aboveq]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  RET
+
+INIT_MMX sse
+cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m1
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m1
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m1
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m1
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_MMX sse
+cglobal tm_predictor_4x4, 4, 4, 4, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  movd                  m0, [aboveq]
+  punpcklbw             m2, m1
+  punpcklbw             m0, m1
+  pshufw                m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -2
+  add                leftq, 4
+  psubw                 m0, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  movd                  m3, [leftq+lineq*2+1]
+  punpcklbw             m2, m1
+  punpcklbw             m3, m1
+  pshufw                m2, m2, 0x0
+  pshufw                m3, m3, 0x0
+  paddw                 m2, m0
+  paddw                 m3, m0
+  packuswb              m2, m2
+  packuswb              m3, m3
+  movd      [dstq        ], m2
+  movd      [dstq+strideq], m3
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_8x8, 4, 4, 4, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  movq                  m0, [aboveq]
+  punpcklbw             m2, m1
+  punpcklbw             m0, m1
+  pshuflw               m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -4
+  punpcklqdq            m2, m2
+  add                leftq, 8
+  psubw                 m0, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  movd                  m3, [leftq+lineq*2+1]
+  punpcklbw             m2, m1
+  punpcklbw             m3, m1
+  pshuflw               m2, m2, 0x0
+  pshuflw               m3, m3, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m3, m3
+  paddw                 m2, m0
+  paddw                 m3, m0
+  packuswb              m2, m3
+  movq      [dstq        ], m2
+  movhps    [dstq+strideq], m2
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_16x16, 4, 4, 7, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  mova                  m0, [aboveq]
+  punpcklbw             m2, m1
+  punpckhbw             m4, m0, m1
+  punpcklbw             m0, m1
+  pshuflw               m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -8
+  punpcklqdq            m2, m2
+  add                leftq, 16
+  psubw                 m0, m2
+  psubw                 m4, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  movd                  m3, [leftq+lineq*2+1]
+  punpcklbw             m2, m1
+  punpcklbw             m3, m1
+  pshuflw               m2, m2, 0x0
+  pshuflw               m3, m3, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m3, m3
+  paddw                 m5, m2, m0
+  paddw                 m6, m3, m0
+  paddw                 m2, m4
+  paddw                 m3, m4
+  packuswb              m5, m2
+  packuswb              m6, m3
+  mova      [dstq        ], m5
+  mova      [dstq+strideq], m6
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+%if ARCH_X86_64
+INIT_XMM sse2
+cglobal tm_predictor_32x32, 4, 4, 10, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  mova                  m0, [aboveq]
+  mova                  m4, [aboveq+16]
+  punpcklbw             m2, m1
+  punpckhbw             m3, m0, m1
+  punpckhbw             m5, m4, m1
+  punpcklbw             m0, m1
+  punpcklbw             m4, m1
+  pshuflw               m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -16
+  punpcklqdq            m2, m2
+  add                leftq, 32
+  psubw                 m0, m2
+  psubw                 m3, m2
+  psubw                 m4, m2
+  psubw                 m5, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  movd                  m6, [leftq+lineq*2+1]
+  punpcklbw             m2, m1
+  punpcklbw             m6, m1
+  pshuflw               m2, m2, 0x0
+  pshuflw               m6, m6, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m6, m6
+  paddw                 m7, m2, m0
+  paddw                 m8, m2, m3
+  paddw                 m9, m2, m4
+  paddw                 m2, m5
+  packuswb              m7, m8
+  packuswb              m9, m2
+  paddw                 m2, m6, m0
+  paddw                 m8, m6, m3
+  mova   [dstq           ], m7
+  paddw                 m7, m6, m4
+  paddw                 m6, m5
+  mova   [dstq        +16], m9
+  packuswb              m2, m8
+  packuswb              m7, m6
+  mova   [dstq+strideq   ], m2
+  mova   [dstq+strideq+16], m7
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+%endif
diff --git a/media/libvpx/vp9/common/x86/vp9_intrapred_ssse3.asm b/media/libvpx/vp9/common/x86/vp9_intrapred_ssse3.asm
new file mode 100644
index 000000000..88df9b2d1
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_intrapred_ssse3.asm
@@ -0,0 +1,1036 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+
+pb_1: times 16 db 1
+sh_b01234577: db 0, 1, 2, 3, 4, 5, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b23456777: db 2, 3, 4, 5, 6, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b0123456777777777: db 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b1234567777777777: db 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b2345677777777777: db 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b123456789abcdeff: db 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15
+sh_b23456789abcdefff: db 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15, 15
+sh_b32104567: db 3, 2, 1, 0, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b8091a2b345: db 8, 0, 9, 1, 10, 2, 11, 3, 4, 5, 0, 0, 0, 0, 0, 0
+sh_b76543210: db 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b65432108: db 6, 5, 4, 3, 2, 1, 0, 8, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b54321089: db 5, 4, 3, 2, 1, 0, 8, 9, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b89abcdef: db 8, 9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0
+sh_bfedcba9876543210: db 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
+sh_b1233: db 1, 2, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b2333: db 2, 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+
+SECTION .text
+
+INIT_MMX ssse3
+cglobal h_predictor_4x4, 2, 4, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  add                leftq, 4
+  mov                lineq, -2
+  pxor                  m0, m0
+.loop:
+  movd                  m1, [leftq+lineq*2  ]
+  movd                  m2, [leftq+lineq*2+1]
+  pshufb                m1, m0
+  pshufb                m2, m0
+  movd      [dstq        ], m1
+  movd      [dstq+strideq], m2
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_MMX ssse3
+cglobal h_predictor_8x8, 2, 4, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  add                leftq, 8
+  mov                lineq, -4
+  pxor                  m0, m0
+.loop:
+  movd                  m1, [leftq+lineq*2  ]
+  movd                  m2, [leftq+lineq*2+1]
+  pshufb                m1, m0
+  pshufb                m2, m0
+  movq      [dstq        ], m1
+  movq      [dstq+strideq], m2
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM ssse3
+cglobal h_predictor_16x16, 2, 4, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  add                leftq, 16
+  mov                lineq, -8
+  pxor                  m0, m0
+.loop:
+  movd                  m1, [leftq+lineq*2  ]
+  movd                  m2, [leftq+lineq*2+1]
+  pshufb                m1, m0
+  pshufb                m2, m0
+  mova      [dstq        ], m1
+  mova      [dstq+strideq], m2
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM ssse3
+cglobal h_predictor_32x32, 2, 4, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  add                leftq, 32
+  mov                lineq, -16
+  pxor                  m0, m0
+.loop:
+  movd                  m1, [leftq+lineq*2  ]
+  movd                  m2, [leftq+lineq*2+1]
+  pshufb                m1, m0
+  pshufb                m2, m0
+  mova   [dstq           ], m1
+  mova   [dstq        +16], m1
+  mova   [dstq+strideq   ], m2
+  mova   [dstq+strideq+16], m2
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_MMX ssse3
+cglobal d45_predictor_4x4, 3, 4, 4, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m0, [aboveq]
+  pshufb              m2, m0, [GLOBAL(sh_b23456777)]
+  pshufb              m1, m0, [GLOBAL(sh_b01234577)]
+  pshufb              m0, [GLOBAL(sh_b12345677)]
+  pavgb               m3, m2, m1
+  pxor                m2, m1
+  pand                m2, [GLOBAL(pb_1)]
+  psubb               m3, m2
+  pavgb               m0, m3
+
+  ; store 4 lines
+  movd    [dstq        ], m0
+  psrlq               m0, 8
+  movd    [dstq+strideq], m0
+  lea               dstq, [dstq+strideq*2]
+  psrlq               m0, 8
+  movd    [dstq        ], m0
+  psrlq               m0, 8
+  movd    [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX ssse3
+cglobal d45_predictor_8x8, 3, 4, 4, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m0, [aboveq]
+  mova                m1, [GLOBAL(sh_b12345677)]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m2, m0, [GLOBAL(sh_b23456777)]
+  pavgb               m3, m2, m0
+  pxor                m2, m0
+  pshufb              m0, m1
+  pand                m2, [GLOBAL(pb_1)]
+  psubb               m3, m2
+  pavgb               m0, m3
+
+  ; store 4 lines
+  movq  [dstq          ], m0
+  pshufb              m0, m1
+  movq  [dstq+strideq  ], m0
+  pshufb              m0, m1
+  movq  [dstq+strideq*2], m0
+  pshufb              m0, m1
+  movq  [dstq+stride3q ], m0
+  pshufb              m0, m1
+  lea               dstq, [dstq+strideq*4]
+
+  ; store next 4 lines
+  movq  [dstq          ], m0
+  pshufb              m0, m1
+  movq  [dstq+strideq  ], m0
+  pshufb              m0, m1
+  movq  [dstq+strideq*2], m0
+  pshufb              m0, m1
+  movq  [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d45_predictor_16x16, 3, 6, 4, dst, stride, above, dst8, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, dst8, line
+  lea              stride3q, [strideq*3]
+  lea                 dst8q, [dstq+strideq*8]
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                 m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+  pavgb                  m3, m2, m0
+  pxor                   m2, m0
+  pshufb                 m0, m1
+  pand                   m2, [GLOBAL(pb_1)]
+  psubb                  m3, m2
+  pavgb                  m0, m3
+
+  ; first 4 lines and first half of 3rd 4 lines
+  mov                 lined, 2
+.loop:
+  mova   [dstq            ], m0
+  movhps [dst8q           ], m0
+  pshufb                 m0, m1
+  mova   [dstq +strideq   ], m0
+  movhps [dst8q+strideq   ], m0
+  pshufb                 m0, m1
+  mova   [dstq +strideq*2 ], m0
+  movhps [dst8q+strideq*2 ], m0
+  pshufb                 m0, m1
+  mova   [dstq +stride3q  ], m0
+  movhps [dst8q+stride3q  ], m0
+  pshufb                 m0, m1
+  lea                  dstq, [dstq +strideq*4]
+  lea                 dst8q, [dst8q+strideq*4]
+  dec                 lined
+  jnz .loop
+
+  ; bottom-right 8x8 block
+  movhps [dstq          +8], m0
+  movhps [dstq+strideq  +8], m0
+  movhps [dstq+strideq*2+8], m0
+  movhps [dstq+stride3q +8], m0
+  lea                  dstq, [dstq+strideq*4]
+  movhps [dstq          +8], m0
+  movhps [dstq+strideq  +8], m0
+  movhps [dstq+strideq*2+8], m0
+  movhps [dstq+stride3q +8], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d45_predictor_32x32, 3, 6, 7, dst, stride, above, dst16, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  mova                   m4, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, dst16, line
+  lea              stride3q, [strideq*3]
+  lea                dst16q, [dstq  +strideq*8]
+  lea                dst16q, [dst16q+strideq*8]
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                 m2, m4, [GLOBAL(sh_b23456789abcdefff)]
+  pavgb                  m3, m2, m4
+  pxor                   m2, m4
+  palignr                m5, m4, m0, 1
+  palignr                m6, m4, m0, 2
+  pshufb                 m4, m1
+  pand                   m2, [GLOBAL(pb_1)]
+  psubb                  m3, m2
+  pavgb                  m4, m3
+  pavgb                  m3, m0, m6
+  pxor                   m0, m6
+  pand                   m0, [GLOBAL(pb_1)]
+  psubb                  m3, m0
+  pavgb                  m5, m3
+
+  ; write 4x4 lines (and the first half of the second 4x4 lines)
+  mov                  lined, 4
+.loop:
+  mova [dstq               ], m5
+  mova [dstq            +16], m4
+  mova [dst16q             ], m4
+  palignr                 m3, m4, m5, 1
+  pshufb                  m4, m1
+  mova [dstq  +strideq     ], m3
+  mova [dstq  +strideq  +16], m4
+  mova [dst16q+strideq     ], m4
+  palignr                 m5, m4, m3, 1
+  pshufb                  m4, m1
+  mova [dstq  +strideq*2   ], m5
+  mova [dstq  +strideq*2+16], m4
+  mova [dst16q+strideq*2   ], m4
+  palignr                 m3, m4, m5, 1
+  pshufb                  m4, m1
+  mova [dstq  +stride3q    ], m3
+  mova [dstq  +stride3q +16], m4
+  mova [dst16q+stride3q    ], m4
+  palignr                 m5, m4, m3, 1
+  pshufb                  m4, m1
+  lea                  dstq, [dstq  +strideq*4]
+  lea                dst16q, [dst16q+strideq*4]
+  dec                 lined
+  jnz .loop
+
+  ; write second half of second 4x4 lines
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+  lea                  dstq, [dstq  +strideq*4]
+  mova [dstq            +16], m4
+  mova [dstq  +strideq  +16], m4
+  mova [dstq  +strideq*2+16], m4
+  mova [dstq  +stride3q +16], m4
+
+  RESTORE_GOT
+  RET
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+  pavgb               %4, %1, %3
+  pxor                %3, %1
+  pand                %3, [GLOBAL(pb_1)]
+  psubb               %4, %3
+  pavgb               %4, %2
+%endmacro
+
+INIT_XMM ssse3
+cglobal d63_predictor_4x4, 3, 4, 5, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m3, [aboveq]
+  pshufb              m1, m3, [GLOBAL(sh_b23456777)]
+  pshufb              m2, m3, [GLOBAL(sh_b12345677)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m2, m1, m4
+  pavgb               m3, m2
+
+  ; store 4 lines
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  lea               dstq, [dstq+strideq*2]
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_8x8, 3, 4, 5, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m3, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m1, m3, [GLOBAL(sh_b2345677777777777)]
+  pshufb              m0, m3, [GLOBAL(sh_b0123456777777777)]
+  pshufb              m2, m3, [GLOBAL(sh_b1234567777777777)]
+  pshufb              m3, [GLOBAL(sh_b0123456777777777)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m2, m1, m4
+  pavgb               m3, m2
+
+  ; store 4 lines
+  movq    [dstq        ], m3
+  movq    [dstq+strideq], m4
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movq  [dstq+strideq*2], m3
+  movq  [dstq+stride3q ], m4
+  lea               dstq, [dstq+strideq*4]
+  psrldq              m3, 1
+  psrldq              m4, 1
+
+  ; store 4 lines
+  movq    [dstq        ], m3
+  movq    [dstq+strideq], m4
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movq  [dstq+strideq*2], m3
+  movq  [dstq+stride3q ], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_16x16, 3, 5, 5, dst, stride, above, line, goffset
+  GET_GOT     goffsetq
+
+  mova                m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, line
+  lea           stride3q, [strideq*3]
+  mova                m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb              m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb              m3, m0, m1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m3, m2, m4
+  pavgb               m0, m3
+
+  mov              lined, 4
+.loop:
+  mova  [dstq          ], m0
+  mova  [dstq+strideq  ], m4
+  pshufb              m0, m1
+  pshufb              m4, m1
+  mova  [dstq+strideq*2], m0
+  mova  [dstq+stride3q ], m4
+  pshufb              m0, m1
+  pshufb              m4, m1
+  lea               dstq, [dstq+strideq*4]
+  dec              lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d63_predictor_32x32, 3, 5, 8, dst, stride, above, line, goffset
+  GET_GOT     goffsetq
+
+  mova                   m0, [aboveq]
+  mova                   m7, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, line
+  mova                   m1, [GLOBAL(sh_b123456789abcdeff)]
+  lea              stride3q, [strideq*3]
+  pshufb                 m2, m7, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb                 m3, m7, m1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m2, m4
+  palignr                m6, m7, m0, 1
+  palignr                m5, m7, m0, 2
+  pavgb                  m7, m3
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m6, m5, m2
+  pavgb                  m0, m6
+
+  mov                 lined, 8
+.loop:
+  mova  [dstq             ], m0
+  mova  [dstq          +16], m7
+  mova  [dstq+strideq     ], m2
+  mova  [dstq+strideq  +16], m4
+  palignr                m3, m7, m0, 1
+  palignr                m5, m4, m2, 1
+  pshufb                 m7, m1
+  pshufb                 m4, m1
+
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m5
+  mova  [dstq+stride3q +16], m4
+  palignr                m0, m7, m3, 1
+  palignr                m2, m4, m5, 1
+  pshufb                 m7, m1
+  pshufb                 m4, m1
+  lea                  dstq, [dstq+strideq*4]
+  dec                 lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movd                m0, [leftq]               ; l1, l2, l3, l4
+  movd                m1, [aboveq-1]            ; tl, t1, t2, t3
+  punpckldq           m0, m1                    ; l1, l2, l3, l4, tl, t1, t2, t3
+  pshufb              m0, [GLOBAL(sh_b32104567)]; l4, l3, l2, l1, tl, t1, t2, t3
+  psrldq              m1, m0, 1                 ; l3, l2, l1, tl, t1, t2, t3
+  psrldq              m2, m0, 2                 ; l2, l1, tl, t1, t2, t3
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1
+  ; A2 B2 A1 B1
+  ; A3 B3 A2 B2
+  ; A4 B4 A3 B3
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3  ; 3-tap avg B4 B3 B2 B1 C1 D1
+  pavgb               m1, m0                    ; 2-tap avg A4 A3 A2 A1
+
+  punpcklqdq          m3, m1                    ; B4 B3 B2 B1 C1 D1 x x A4 A3 A2 A1 ..
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m3, [GLOBAL(sh_b8091a2b345)] ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+stride3q ], m3
+  psrldq              m3, 2                     ; A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq*2], m3
+  psrldq              m3, 2                     ; A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq  ], m3
+  psrldq              m3, 2                     ; A1 B1 C1 D1 ..
+  movd  [dstq          ], m3
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_8x8, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movq                m0, [leftq]                     ; [0- 7] l1-8 [byte]
+  movhps              m0, [aboveq-1]                  ; [8-15] tl, t1-7 [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b76543210)]  ; l8-1 [word]
+  pshufb              m2, m0, [GLOBAL(sh_b65432108)]  ; l7-1,tl [word]
+  pshufb              m3, m0, [GLOBAL(sh_b54321089)]  ; l6-1,tl,t1 [word]
+  pshufb              m0, [GLOBAL(sh_b89abcdef)]      ; tl,t1-7 [word]
+  psrldq              m4, m0, 1                       ; t1-7 [word]
+  psrldq              m5, m0, 2                       ; t2-7 [word]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1
+  ; A2 B2 A1 B1 C1 D1 E1 F1
+  ; A3 B3 A2 B2 A1 B1 C1 D1
+  ; A4 B4 A3 B3 A2 B2 A1 B1
+  ; A5 B5 A4 B4 A3 B3 A2 B2
+  ; A6 B6 A5 B5 A4 B4 A3 B3
+  ; A7 B7 A6 B6 A5 B5 A4 B4
+  ; A8 B8 A7 B7 A6 B6 A5 B5
+  pavgb               m6, m1, m2                ; 2-tap avg A8-A1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m4, m5, m7  ; 3-tap avg C-H1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m2, m3, m0  ; 3-tap avg B8-1
+
+  punpcklbw           m6, m0                    ; A-B8, A-B7 ... A-B2, A-B1
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+
+  movhps [dstq+stride3q], m6                    ; A-B4, A-B3, A-B2, A-B1
+  palignr             m0, m7, m6, 10            ; A-B3, A-B2, A-B1, C-H1
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2                     ; A-B2, A-B1, C-H1
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2                     ; A-H1
+  movq  [dstq          ], m0
+  lea               dstq, [dstq+strideq*4]
+  movq  [dstq+stride3q ], m6                    ; A-B8, A-B7, A-B6, A-B5
+  psrldq              m6, 2                     ; A-B7, A-B6, A-B5, A-B4
+  movq  [dstq+strideq*2], m6
+  psrldq              m6, 2                     ; A-B6, A-B5, A-B4, A-B3
+  movq  [dstq+strideq  ], m6
+  psrldq              m6, 2                     ; A-B5, A-B4, A-B3, A-B2
+  movq  [dstq          ], m6
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_16x16, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                m0, [leftq]
+  movu                m7, [aboveq-1]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1 O1 P1
+  ; A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1
+  ; A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1
+  ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1
+  ; A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1
+  ; A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1
+  ; A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1
+  ; A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1
+  ; A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2
+  ; Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3
+  ; Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4
+  ; Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5
+  ; Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6
+  ; Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7
+  ; Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8
+  ; Ag Bg Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9
+  pshufb              m6, m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr             m5, m0, m6, 15
+  palignr             m3, m0, m6, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pshufb              m1, m0, [GLOBAL(sh_b123456789abcdeff)]
+  pavgb               m5, m0                            ; A1 - Ag
+
+  punpcklbw           m0, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw           m4, m5                            ; A-B9 ... A-Bg
+
+  pshufb              m3, m7, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb              m5, m7, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg C1-P1
+
+  pshufb              m6, m0, [GLOBAL(sh_bfedcba9876543210)]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  palignr             m2, m1, m6, 14
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m1, m6, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m1, m6, 6
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 2
+  mova  [dstq+strideq*2], m2
+  pshufb              m4, [GLOBAL(sh_bfedcba9876543210)]
+  mova  [dstq+stride3q ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  palignr             m2, m6, m4, 14
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m6, m4, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m6, m4, 6
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 2
+  mova  [dstq+strideq*2], m2
+  mova  [dstq+stride3q ], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_32x32, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                  m0, [leftq]
+  movu                  m7, [aboveq-1]
+  movu                  m1, [aboveq+15]
+
+  pshufb                m4, m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                m6, m1, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m4, m6, m2          ; 3-tap avg above [high]
+
+  palignr               m3, m1, m7, 1
+  palignr               m5, m1, m7, 2
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg above [low]
+
+  pshufb                m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr               m5, m0, m7, 15
+  palignr               m3, m0, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pavgb                 m5, m0                            ; A1 - Ag
+  punpcklbw             m6, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw             m4, m5                            ; A-B9 ... A-Bg
+  pshufb                m6, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                m4, [GLOBAL(sh_bfedcba9876543210)]
+
+  DEFINE_ARGS dst, stride, stride3, left, line
+  lea             stride3q, [strideq*3]
+
+  palignr               m5, m2, m1, 14
+  palignr               m7, m1, m6, 14
+  mova  [dstq            ], m7
+  mova  [dstq+16         ], m5
+  palignr               m5, m2, m1, 12
+  palignr               m7, m1, m6, 12
+  mova  [dstq+strideq    ], m7
+  mova  [dstq+strideq+16 ], m5
+  palignr                m5, m2, m1, 10
+  palignr                m7, m1, m6, 10
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m2, m1, 8
+  palignr                m7, m1, m6, 8
+  mova  [dstq+stride3q    ], m7
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m2, m1, 6
+  palignr                m7, m1, m6, 6
+  mova  [dstq             ], m7
+  mova  [dstq+16          ], m5
+  palignr                m5, m2, m1, 4
+  palignr                m7, m1, m6, 4
+  mova  [dstq+strideq     ], m7
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m2, m1, 2
+  palignr                m7, m1, m6, 2
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m6
+  mova  [dstq+stride3q+16 ], m1
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m5, m1, m6, 14
+  palignr                m3, m6, m4, 14
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 12
+  palignr                m3, m6, m4, 12
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 10
+  palignr                m3, m6, m4, 10
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m1, m6, 8
+  palignr                m3, m6, m4, 8
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m1, m6, 6
+  palignr                m3, m6, m4, 6
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 4
+  palignr                m3, m6, m4, 4
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 2
+  palignr                m3, m6, m4, 2
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m4
+  mova  [dstq+stride3q+16 ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  mova                   m7, [leftq]
+  mova                   m3, [leftq+16]
+  palignr                m5, m3, m7, 15
+  palignr                m0, m3, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m5, m0, m2          ; 3-tap avg Bh -
+  pavgb                  m5, m3                            ; Ah -
+  punpcklbw              m3, m2, m5                        ; A-B8 ... A-B1
+  punpckhbw              m2, m5                            ; A-B9 ... A-Bg
+  pshufb                 m3, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                 m2, [GLOBAL(sh_bfedcba9876543210)]
+
+  palignr                m7, m6, m4, 14
+  palignr                m0, m4, m3, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 12
+  palignr                m0, m4, m3, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 10
+  palignr                m0, m4, m3, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m6, m4, 8
+  palignr                m0, m4, m3, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m6, m4, 6
+  palignr                m0, m4, m3, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 4
+  palignr                m0, m4, m3, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 2
+  palignr                m0, m4, m3, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m4
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m7, m4, m3, 14
+  palignr                m0, m3, m2, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 12
+  palignr                m0, m3, m2, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 10
+  palignr                m0, m3, m2, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m4, m3, 8
+  palignr                m0, m3, m2, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m4, m3, 6
+  palignr                m0, m3, m2, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 4
+  palignr                m0, m3, m2, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 2
+  palignr                m0, m3, m2, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m2
+  mova  [dstq+stride3q+16 ], m3
+
+  RESTORE_GOT
+  RET
+
+INIT_MMX ssse3
+cglobal d207_predictor_4x4, 4, 5, 4, dst, stride, unused, left, goffset
+  GET_GOT     goffsetq
+  movd                m0, [leftq]                ; abcd [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b1233)] ; bcdd [byte]
+  pshufb              m3, m0, [GLOBAL(sh_b2333)] ; cddd
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m3, m2
+  pavgb               m1, m0             ; ab, bc, cd, d [byte]
+
+  punpcklbw           m1, m2             ; ab, a2bc, bc, b2cd, cd, c3d, d, d
+  movd    [dstq        ], m1
+  psrlq               m1, 16             ; bc, b2cd, cd, c3d, d, d
+  movd    [dstq+strideq], m1
+  lea               dstq, [dstq+strideq*2]
+  psrlq               m1, 16             ; cd, c3d, d, d
+  movd    [dstq        ], m1
+  pshufw              m1, m1, q1111      ; d, d, d, d
+  movd    [dstq+strideq], m1
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_8x8, 4, 5, 4, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  movq                m3, [leftq]            ; abcdefgh [byte]
+  lea           stride3q, [strideq*3]
+
+  pshufb              m1, m3, [GLOBAL(sh_b2345677777777777)]
+  pshufb              m0, m3, [GLOBAL(sh_b0123456777777777)]
+  pshufb              m2, m3, [GLOBAL(sh_b1234567777777777)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m2, m1, m3
+  pavgb               m0, m2
+  punpcklbw           m0, m3        ; interleaved output
+
+  movq  [dstq          ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2
+  movq  [dstq+stride3q ], m0
+  lea               dstq, [dstq+strideq*4]
+  pshufhw             m0, m0, q0000 ; de, d2ef, ef, e2fg, fg, f2gh, gh, g3h, 8xh
+  psrldq              m0, 2
+  movq  [dstq          ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2
+  movq  [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_16x16, 4, 5, 5, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  lea           stride3q, [strideq*3]
+  mova                m0, [leftq]            ; abcdefghijklmnop [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b123456789abcdeff)] ; bcdefghijklmnopp
+  pshufb              m2, m0, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3
+  pavgb               m1, m0                 ; ab, bc, cd .. no, op, pp [byte]
+
+  punpckhbw           m4, m1, m3    ; interleaved input
+  punpcklbw           m1, m3        ; interleaved output
+  mova  [dstq          ], m1
+  palignr             m3, m4, m1, 2
+  mova  [dstq+strideq  ], m3
+  palignr             m3, m4, m1, 4
+  mova  [dstq+strideq*2], m3
+  palignr             m3, m4, m1, 6
+  mova  [dstq+stride3q ], m3
+  lea               dstq, [dstq+strideq*4]
+  palignr             m3, m4, m1, 8
+  mova  [dstq          ], m3
+  palignr             m3, m4, m1, 10
+  mova  [dstq+strideq  ], m3
+  palignr             m3, m4, m1, 12
+  mova  [dstq+strideq*2], m3
+  palignr             m3, m4, m1, 14
+  mova  [dstq+stride3q ], m3
+  DEFINE_ARGS dst, stride, stride3, line
+  mov              lined, 2
+  mova                m0, [GLOBAL(sh_b23456789abcdefff)]
+.loop:
+  lea               dstq, [dstq+strideq*4]
+  mova  [dstq          ], m4
+  pshufb              m4, m0
+  mova  [dstq+strideq  ], m4
+  pshufb              m4, m0
+  mova  [dstq+strideq*2], m4
+  pshufb              m4, m0
+  mova  [dstq+stride3q ], m4
+  pshufb              m4, m0
+  dec              lined
+  jnz .loop
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM ssse3
+cglobal d207_predictor_32x32, 4, 5, 8, dst, stride, stride3, left, goffset
+  GET_GOT     goffsetq
+  lea           stride3q, [strideq*3]
+  mova                m1, [leftq]              ;  0-15 [byte]
+  mova                m2, [leftq+16]           ; 16-31 [byte]
+  pshufb              m0, m2, [GLOBAL(sh_b23456789abcdefff)]
+  pshufb              m4, m2, [GLOBAL(sh_b123456789abcdeff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m2, m4, m0, m3
+  palignr             m6, m2, m1, 1
+  palignr             m5, m2, m1, 2
+  pavgb               m2, m4         ; high 16px even lines
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m6, m5, m0
+  pavgb                   m1, m6         ; low 16px even lines
+
+  punpckhbw               m6, m1, m0               ; interleaved output 2
+  punpcklbw               m1, m0                   ; interleaved output 1
+
+  punpckhbw               m7, m2, m3               ; interleaved output 4
+  punpcklbw               m2, m3                   ; interleaved output 3
+
+  ; output 1st 8 lines (and half of 2nd 8 lines)
+  DEFINE_ARGS dst, stride, stride3, dst8
+  lea                  dst8q, [dstq+strideq*8]
+  mova  [dstq              ], m1
+  mova  [dstq           +16], m6
+  mova  [dst8q             ], m6
+  palignr             m0, m6, m1, 2
+  palignr             m4, m2, m6, 2
+  mova  [dstq +strideq     ], m0
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m0, m6, m1, 4
+  palignr             m4, m2, m6, 4
+  mova  [dstq +strideq*2   ], m0
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m0, m6, m1, 6
+  palignr             m4, m2, m6, 6
+  mova  [dstq +stride3q    ], m0
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq +strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  palignr             m0, m6, m1, 8
+  palignr             m4, m2, m6, 8
+  mova  [dstq              ], m0
+  mova  [dstq           +16], m4
+  mova  [dst8q             ], m4
+  palignr             m0, m6, m1, 10
+  palignr             m4, m2, m6, 10
+  mova  [dstq +strideq     ], m0
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m0, m6, m1, 12
+  palignr             m4, m2, m6, 12
+  mova  [dstq +strideq*2   ], m0
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m0, m6, m1, 14
+  palignr             m4, m2, m6, 14
+  mova  [dstq +stride3q    ], m0
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+
+  ; output 2nd half of 2nd 8 lines and half of 3rd 8 lines
+  mova  [dstq           +16], m2
+  mova  [dst8q             ], m2
+  palignr             m4, m7, m2, 2
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m4, m7, m2, 4
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m4, m7, m2, 6
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  palignr             m4, m7, m2, 8
+  mova  [dstq           +16], m4
+  mova  [dst8q             ], m4
+  palignr             m4, m7, m2, 10
+  mova  [dstq +strideq  +16], m4
+  mova  [dst8q+strideq     ], m4
+  palignr             m4, m7, m2, 12
+  mova  [dstq +strideq*2+16], m4
+  mova  [dst8q+strideq*2   ], m4
+  palignr             m4, m7, m2, 14
+  mova  [dstq +stride3q +16], m4
+  mova  [dst8q+stride3q    ], m4
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+
+  ; output 2nd half of 3rd 8 lines and half of 4th 8 lines
+  mova                m0, [GLOBAL(sh_b23456789abcdefff)]
+  mova  [dstq           +16], m7
+  mova  [dst8q             ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq  +16], m7
+  mova  [dst8q+strideq     ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq*2+16], m7
+  mova  [dst8q+strideq*2   ], m7
+  pshufb              m7, m0
+  mova  [dstq +stride3q +16], m7
+  mova  [dst8q+stride3q    ], m7
+  pshufb              m7, m0
+  lea               dstq, [dstq+strideq*4]
+  lea              dst8q, [dst8q+strideq*4]
+  mova  [dstq           +16], m7
+  mova  [dst8q             ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq  +16], m7
+  mova  [dst8q+strideq     ], m7
+  pshufb              m7, m0
+  mova  [dstq +strideq*2+16], m7
+  mova  [dst8q+strideq*2   ], m7
+  pshufb              m7, m0
+  mova  [dstq +stride3q +16], m7
+  mova  [dst8q+stride3q    ], m7
+  pshufb              m7, m0
+  lea               dstq, [dstq+strideq*4]
+
+  ; output last half of 4th 8 lines
+  mova  [dstq           +16], m7
+  mova  [dstq +strideq  +16], m7
+  mova  [dstq +strideq*2+16], m7
+  mova  [dstq +stride3q +16], m7
+  lea               dstq, [dstq+strideq*4]
+  mova  [dstq           +16], m7
+  mova  [dstq +strideq  +16], m7
+  mova  [dstq +strideq*2+16], m7
+  mova  [dstq +stride3q +16], m7
+
+  ; done!
+  RESTORE_GOT
+  RET
diff --git a/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_avx2.c b/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
new file mode 100644
index 000000000..770a65f4c
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
@@ -0,0 +1,986 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  /* AVX2 */
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+
+static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
+        const unsigned char *_blimit, const unsigned char *_limit,
+        const unsigned char *_thresh) {
+    __m128i mask, hev, flat, flat2;
+    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+    __m128i abs_p1p0;
+
+    const __m128i thresh = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _thresh[0]));
+    const __m128i limit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _limit[0]));
+    const __m128i blimit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _blimit[0]));
+
+    q4p4 = _mm_loadl_epi64((__m128i *) (s - 5 * p));
+    q4p4 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *) (s + 4 * p)));
+    q3p3 = _mm_loadl_epi64((__m128i *) (s - 4 * p));
+    q3p3 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *) (s + 3 * p)));
+    q2p2 = _mm_loadl_epi64((__m128i *) (s - 3 * p));
+    q2p2 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *) (s + 2 * p)));
+    q1p1 = _mm_loadl_epi64((__m128i *) (s - 2 * p));
+    q1p1 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *) (s + 1 * p)));
+    p1q1 = _mm_shuffle_epi32(q1p1, 78);
+    q0p0 = _mm_loadl_epi64((__m128i *) (s - 1 * p));
+    q0p0 = _mm_castps_si128(
+            _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *) (s - 0 * p)));
+    p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+    {
+        __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+        abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
+                _mm_subs_epu8(q0p0, q1p1));
+        abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+        fe = _mm_set1_epi8(0xfe);
+        ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+        abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
+                _mm_subs_epu8(p0q0, q0p0));
+        abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
+                _mm_subs_epu8(p1q1, q1p1));
+        flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+        hev = _mm_subs_epu8(flat, thresh);
+        hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+        abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+        abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+        mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+        mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+        // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+        mask = _mm_max_epu8(abs_p1p0, mask);
+        // mask |= (abs(p1 - p0) > limit) * -1;
+        // mask |= (abs(q1 - q0) > limit) * -1;
+
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
+                        _mm_subs_epu8(q1p1, q2p2)),
+                _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
+                        _mm_subs_epu8(q2p2, q3p3)));
+        mask = _mm_max_epu8(work, mask);
+        mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+        mask = _mm_subs_epu8(mask, limit);
+        mask = _mm_cmpeq_epi8(mask, zero);
+    }
+
+    // lp filter
+    {
+        const __m128i t4 = _mm_set1_epi8(4);
+        const __m128i t3 = _mm_set1_epi8(3);
+        const __m128i t80 = _mm_set1_epi8(0x80);
+        const __m128i t1 = _mm_set1_epi16(0x1);
+        __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+        __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+        __m128i qs0 = _mm_xor_si128(p0q0, t80);
+        __m128i qs1 = _mm_xor_si128(p1q1, t80);
+        __m128i filt;
+        __m128i work_a;
+        __m128i filter1, filter2;
+        __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+        __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+        filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+        work_a = _mm_subs_epi8(qs0, qs0ps0);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
+        filt = _mm_and_si128(filt, mask);
+
+        filter1 = _mm_adds_epi8(filt, t4);
+        filter2 = _mm_adds_epi8(filt, t3);
+
+        filter1 = _mm_unpacklo_epi8(zero, filter1);
+        filter1 = _mm_srai_epi16(filter1, 0xB);
+        filter2 = _mm_unpacklo_epi8(zero, filter2);
+        filter2 = _mm_srai_epi16(filter2, 0xB);
+
+        /* Filter1 >> 3 */
+        filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+        qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+        /* filt >> 1 */
+        filt = _mm_adds_epi16(filter1, t1);
+        filt = _mm_srai_epi16(filt, 1);
+        filt = _mm_andnot_si128(
+                _mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), filt);
+        filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+        qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+        // loopfilter done
+
+        {
+            __m128i work;
+            flat = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
+                            _mm_subs_epu8(q0p0, q2p2)),
+                    _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
+                            _mm_subs_epu8(q0p0, q3p3)));
+            flat = _mm_max_epu8(abs_p1p0, flat);
+            flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+            flat = _mm_subs_epu8(flat, one);
+            flat = _mm_cmpeq_epi8(flat, zero);
+            flat = _mm_and_si128(flat, mask);
+
+            q5p5 = _mm_loadl_epi64((__m128i *) (s - 6 * p));
+            q5p5 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q5p5),
+                            (__m64 *) (s + 5 * p)));
+
+            q6p6 = _mm_loadl_epi64((__m128i *) (s - 7 * p));
+            q6p6 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q6p6),
+                            (__m64 *) (s + 6 * p)));
+
+            flat2 = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
+                            _mm_subs_epu8(q0p0, q4p4)),
+                    _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
+                            _mm_subs_epu8(q0p0, q5p5)));
+
+            q7p7 = _mm_loadl_epi64((__m128i *) (s - 8 * p));
+            q7p7 = _mm_castps_si128(
+                    _mm_loadh_pi(_mm_castsi128_ps(q7p7),
+                            (__m64 *) (s + 7 * p)));
+
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
+                            _mm_subs_epu8(q0p0, q6p6)),
+                    _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
+                            _mm_subs_epu8(q0p0, q7p7)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+            flat2 = _mm_subs_epu8(flat2, one);
+            flat2 = _mm_cmpeq_epi8(flat2, zero);
+            flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+        }
+
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+        // flat and wide flat calculations
+        {
+            const __m128i eight = _mm_set1_epi16(8);
+            const __m128i four = _mm_set1_epi16(4);
+            __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+            __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+            __m128i pixelFilter_p, pixelFilter_q;
+            __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+            __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+            p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+            p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+            p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+            p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+            p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+            p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+            p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+            p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+            q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+            q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+            q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+            q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+            q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+            q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+            q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+            q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+            pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                    _mm_add_epi16(p4_16, p3_16));
+            pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                    _mm_add_epi16(q4_16, q3_16));
+
+            pixetFilter_p2p1p0 = _mm_add_epi16(p0_16,
+                    _mm_add_epi16(p2_16, p1_16));
+            pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+            pixetFilter_q2q1q0 = _mm_add_epi16(q0_16,
+                    _mm_add_epi16(q2_16, q1_16));
+            pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+            pixelFilter_p = _mm_add_epi16(eight,
+                    _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+            pixetFilter_p2p1p0 = _mm_add_epi16(four,
+                    _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)),
+                    4);
+            flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(p3_16, p0_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(q3_16, q0_16)), 3);
+
+            flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(p7_16, p7_16);
+            sum_q7 = _mm_add_epi16(q7_16, q7_16);
+            sum_p3 = _mm_add_epi16(p3_16, p3_16);
+            sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)),
+                    4);
+            flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+            pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+            pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(sum_p3, p1_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_q2q1q0,
+                            _mm_add_epi16(sum_q3, q1_16)), 3);
+            flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+            sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)),
+                    4);
+            flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+            pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+            pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_p2p1p0,
+                            _mm_add_epi16(sum_p3, p2_16)), 3);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixetFilter_q2q1q0,
+                            _mm_add_epi16(sum_q3, q2_16)), 3);
+            flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)),
+                    4);
+            flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)),
+                    4);
+            flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)),
+                    4);
+            flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+            sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+            sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+            pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+            pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+            res_p = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)),
+                    4);
+            res_q = _mm_srli_epi16(
+                    _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)),
+                    4);
+            flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+        }
+        // wide flat
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+        flat = _mm_shuffle_epi32(flat, 68);
+        flat2 = _mm_shuffle_epi32(flat2, 68);
+
+        q2p2 = _mm_andnot_si128(flat, q2p2);
+        flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+        q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+        qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+        flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+        q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+        qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+        flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+        q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+        q6p6 = _mm_andnot_si128(flat2, q6p6);
+        flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+        q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+        _mm_storel_epi64((__m128i *) (s - 7 * p), q6p6);
+        _mm_storeh_pi((__m64 *) (s + 6 * p), _mm_castsi128_ps(q6p6));
+
+        q5p5 = _mm_andnot_si128(flat2, q5p5);
+        flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+        q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+        _mm_storel_epi64((__m128i *) (s - 6 * p), q5p5);
+        _mm_storeh_pi((__m64 *) (s + 5 * p), _mm_castsi128_ps(q5p5));
+
+        q4p4 = _mm_andnot_si128(flat2, q4p4);
+        flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+        q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+        _mm_storel_epi64((__m128i *) (s - 5 * p), q4p4);
+        _mm_storeh_pi((__m64 *) (s + 4 * p), _mm_castsi128_ps(q4p4));
+
+        q3p3 = _mm_andnot_si128(flat2, q3p3);
+        flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+        q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+        _mm_storel_epi64((__m128i *) (s - 4 * p), q3p3);
+        _mm_storeh_pi((__m64 *) (s + 3 * p), _mm_castsi128_ps(q3p3));
+
+        q2p2 = _mm_andnot_si128(flat2, q2p2);
+        flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+        q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+        _mm_storel_epi64((__m128i *) (s - 3 * p), q2p2);
+        _mm_storeh_pi((__m64 *) (s + 2 * p), _mm_castsi128_ps(q2p2));
+
+        q1p1 = _mm_andnot_si128(flat2, q1p1);
+        flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+        q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+        _mm_storel_epi64((__m128i *) (s - 2 * p), q1p1);
+        _mm_storeh_pi((__m64 *) (s + 1 * p), _mm_castsi128_ps(q1p1));
+
+        q0p0 = _mm_andnot_si128(flat2, q0p0);
+        flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+        q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+        _mm_storel_epi64((__m128i *) (s - 1 * p), q0p0);
+        _mm_storeh_pi((__m64 *) (s - 0 * p), _mm_castsi128_ps(q0p0));
+    }
+}
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
+  0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
+  8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
+static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
+        const unsigned char *_blimit, const unsigned char *_limit,
+        const unsigned char *_thresh) {
+    __m128i mask, hev, flat, flat2;
+    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i p7, p6, p5;
+    __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+    __m128i q5, q6, q7;
+    __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
+            q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
+            p256_0, q256_0;
+
+    const __m128i thresh = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _thresh[0]));
+    const __m128i limit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _limit[0]));
+    const __m128i blimit = _mm_broadcastb_epi8(
+            _mm_cvtsi32_si128((int) _blimit[0]));
+
+    p256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 5 * p)));
+    p256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 4 * p)));
+    p256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 3 * p)));
+    p256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 2 * p)));
+    p256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 1 * p)));
+    q256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s - 0 * p)));
+    q256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 1 * p)));
+    q256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 2 * p)));
+    q256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 3 * p)));
+    q256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                (__m128d const *)(s + 4 * p)));
+
+    p4 = _mm256_castsi256_si128(p256_4);
+    p3 = _mm256_castsi256_si128(p256_3);
+    p2 = _mm256_castsi256_si128(p256_2);
+    p1 = _mm256_castsi256_si128(p256_1);
+    p0 = _mm256_castsi256_si128(p256_0);
+    q0 = _mm256_castsi256_si128(q256_0);
+    q1 = _mm256_castsi256_si128(q256_1);
+    q2 = _mm256_castsi256_si128(q256_2);
+    q3 = _mm256_castsi256_si128(q256_3);
+    q4 = _mm256_castsi256_si128(q256_4);
+
+    {
+        const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                _mm_subs_epu8(p0, p1));
+        const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                _mm_subs_epu8(q0, q1));
+        const __m128i fe = _mm_set1_epi8(0xfe);
+        const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+        __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                _mm_subs_epu8(q0, p0));
+        __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                _mm_subs_epu8(q1, p1));
+        __m128i work;
+        flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+        hev = _mm_subs_epu8(flat, thresh);
+        hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+        abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+        abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+        mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+        mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+        // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+        mask = _mm_max_epu8(flat, mask);
+        // mask |= (abs(p1 - p0) > limit) * -1;
+        // mask |= (abs(q1 - q0) > limit) * -1;
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+                _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+        mask = _mm_max_epu8(work, mask);
+        work = _mm_max_epu8(
+                _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+                _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+        mask = _mm_max_epu8(work, mask);
+        mask = _mm_subs_epu8(mask, limit);
+        mask = _mm_cmpeq_epi8(mask, zero);
+    }
+
+    // lp filter
+    {
+        const __m128i t4 = _mm_set1_epi8(4);
+        const __m128i t3 = _mm_set1_epi8(3);
+        const __m128i t80 = _mm_set1_epi8(0x80);
+        const __m128i te0 = _mm_set1_epi8(0xe0);
+        const __m128i t1f = _mm_set1_epi8(0x1f);
+        const __m128i t1 = _mm_set1_epi8(0x1);
+        const __m128i t7f = _mm_set1_epi8(0x7f);
+
+        __m128i ps1 = _mm_xor_si128(p1, t80);
+        __m128i ps0 = _mm_xor_si128(p0, t80);
+        __m128i qs0 = _mm_xor_si128(q0, t80);
+        __m128i qs1 = _mm_xor_si128(q1, t80);
+        __m128i filt;
+        __m128i work_a;
+        __m128i filter1, filter2;
+        __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
+                flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4,
+                flat2_q5, flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1,
+                flat_q2;
+
+        filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+        work_a = _mm_subs_epi8(qs0, ps0);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        filt = _mm_adds_epi8(filt, work_a);
+        /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
+        filt = _mm_and_si128(filt, mask);
+
+        filter1 = _mm_adds_epi8(filt, t4);
+        filter2 = _mm_adds_epi8(filt, t3);
+
+        /* Filter1 >> 3 */
+        work_a = _mm_cmpgt_epi8(zero, filter1);
+        filter1 = _mm_srli_epi16(filter1, 3);
+        work_a = _mm_and_si128(work_a, te0);
+        filter1 = _mm_and_si128(filter1, t1f);
+        filter1 = _mm_or_si128(filter1, work_a);
+        qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+        /* Filter2 >> 3 */
+        work_a = _mm_cmpgt_epi8(zero, filter2);
+        filter2 = _mm_srli_epi16(filter2, 3);
+        work_a = _mm_and_si128(work_a, te0);
+        filter2 = _mm_and_si128(filter2, t1f);
+        filter2 = _mm_or_si128(filter2, work_a);
+        ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+        /* filt >> 1 */
+        filt = _mm_adds_epi8(filter1, t1);
+        work_a = _mm_cmpgt_epi8(zero, filt);
+        filt = _mm_srli_epi16(filt, 1);
+        work_a = _mm_and_si128(work_a, t80);
+        filt = _mm_and_si128(filt, t7f);
+        filt = _mm_or_si128(filt, work_a);
+        filt = _mm_andnot_si128(hev, filt);
+        ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+        qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+        // loopfilter done
+
+        {
+            __m128i work;
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+                    _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+            flat = _mm_max_epu8(work, flat);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+                    _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+            flat = _mm_max_epu8(work, flat);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
+                    _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
+            flat = _mm_subs_epu8(flat, one);
+            flat = _mm_cmpeq_epi8(flat, zero);
+            flat = _mm_and_si128(flat, mask);
+
+            p256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 6 * p)));
+            q256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 5 * p)));
+            p5 = _mm256_castsi256_si128(p256_5);
+            q5 = _mm256_castsi256_si128(q256_5);
+            flat2 = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
+                    _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            p256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 7 * p)));
+            q256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 6 * p)));
+            p6 = _mm256_castsi256_si128(p256_6);
+            q6 = _mm256_castsi256_si128(q256_6);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
+                    _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+
+            p256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s - 8 * p)));
+            q256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+                                        (__m128d const *)(s + 7 * p)));
+            p7 = _mm256_castsi256_si128(p256_7);
+            q7 = _mm256_castsi256_si128(q256_7);
+            work = _mm_max_epu8(
+                    _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
+                    _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
+
+            flat2 = _mm_max_epu8(work, flat2);
+            flat2 = _mm_subs_epu8(flat2, one);
+            flat2 = _mm_cmpeq_epi8(flat2, zero);
+            flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+        }
+
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+        // flat and wide flat calculations
+        {
+            const __m256i eight = _mm256_set1_epi16(8);
+            const __m256i four = _mm256_set1_epi16(4);
+            __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
+                    pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p,
+                    res_q;
+
+            const __m256i filter = _mm256_load_si256(
+                                  (__m256i const *)filt_loopfilter_avx2);
+            p256_7 = _mm256_shuffle_epi8(p256_7, filter);
+            p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+            p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+            p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+            p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+            p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+            p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+            p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+            q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+            q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+            q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+            q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+            q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+            q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+            q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+            q256_7 = _mm256_shuffle_epi8(q256_7, filter);
+
+            pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
+                    _mm256_add_epi16(p256_4, p256_3));
+            pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
+                    _mm256_add_epi16(q256_4, q256_3));
+
+            pixetFilter_p2p1p0 = _mm256_add_epi16(p256_0,
+                    _mm256_add_epi16(p256_2, p256_1));
+            pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+            pixetFilter_q2q1q0 = _mm256_add_epi16(q256_0,
+                    _mm256_add_epi16(q256_2, q256_1));
+            pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+
+            pixelFilter_p = _mm256_add_epi16(eight,
+                    _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
+
+            pixetFilter_p2p1p0 = _mm256_add_epi16(four,
+                    _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(p256_7, p256_0)), 4);
+
+            flat2_p0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(q256_7, q256_0)), 4);
+
+            flat2_q0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(p256_3, p256_0)), 3);
+
+            flat_p0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(q256_3, q256_0)), 3);
+
+            flat_q0 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(p256_7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(q256_7, q256_7);
+
+            sum_p3 = _mm256_add_epi16(p256_3, p256_3);
+
+            sum_q3 = _mm256_add_epi16(q256_3, q256_3);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_1)), 4);
+
+            flat2_p1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_1)), 4);
+
+            flat2_q1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
+
+            pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(sum_p3, p256_1)), 3);
+
+            flat_p1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_q2q1q0,
+                            _mm256_add_epi16(sum_q3, q256_1)), 3);
+
+            flat_q1 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
+
+            sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_2)), 4);
+
+            flat2_p2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_2)), 4);
+
+            flat2_q2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
+
+            pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_p2p1p0,
+                            _mm256_add_epi16(sum_p3, p256_2)), 3);
+
+            flat_p2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixetFilter_q2q1q0,
+                            _mm256_add_epi16(sum_q3, q256_2)), 3);
+
+            flat_q2 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_3)), 4);
+
+            flat2_p3 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_3)), 4);
+
+            flat2_q3 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_4)), 4);
+
+            flat2_p4 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_4)), 4);
+
+            flat2_q4 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_5)), 4);
+
+            flat2_p5 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_5)), 4);
+
+            flat2_q5 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+
+            sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+            sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+            pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
+
+            pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
+
+            res_p = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_p,
+                            _mm256_add_epi16(sum_p7, p256_6)), 4);
+
+            flat2_p6 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+                            168));
+
+            res_q = _mm256_srli_epi16(
+                    _mm256_add_epi16(pixelFilter_q,
+                            _mm256_add_epi16(sum_q7, q256_6)), 4);
+
+            flat2_q6 = _mm256_castsi256_si128(
+                    _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+                            168));
+        }
+
+        // wide flat
+        // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+        p2 = _mm_andnot_si128(flat, p2);
+        flat_p2 = _mm_and_si128(flat, flat_p2);
+        p2 = _mm_or_si128(flat_p2, p2);
+
+        p1 = _mm_andnot_si128(flat, ps1);
+        flat_p1 = _mm_and_si128(flat, flat_p1);
+        p1 = _mm_or_si128(flat_p1, p1);
+
+        p0 = _mm_andnot_si128(flat, ps0);
+        flat_p0 = _mm_and_si128(flat, flat_p0);
+        p0 = _mm_or_si128(flat_p0, p0);
+
+        q0 = _mm_andnot_si128(flat, qs0);
+        flat_q0 = _mm_and_si128(flat, flat_q0);
+        q0 = _mm_or_si128(flat_q0, q0);
+
+        q1 = _mm_andnot_si128(flat, qs1);
+        flat_q1 = _mm_and_si128(flat, flat_q1);
+        q1 = _mm_or_si128(flat_q1, q1);
+
+        q2 = _mm_andnot_si128(flat, q2);
+        flat_q2 = _mm_and_si128(flat, flat_q2);
+        q2 = _mm_or_si128(flat_q2, q2);
+
+        p6 = _mm_andnot_si128(flat2, p6);
+        flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+        p6 = _mm_or_si128(flat2_p6, p6);
+        _mm_storeu_si128((__m128i *) (s - 7 * p), p6);
+
+        p5 = _mm_andnot_si128(flat2, p5);
+        flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+        p5 = _mm_or_si128(flat2_p5, p5);
+        _mm_storeu_si128((__m128i *) (s - 6 * p), p5);
+
+        p4 = _mm_andnot_si128(flat2, p4);
+        flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+        p4 = _mm_or_si128(flat2_p4, p4);
+        _mm_storeu_si128((__m128i *) (s - 5 * p), p4);
+
+        p3 = _mm_andnot_si128(flat2, p3);
+        flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+        p3 = _mm_or_si128(flat2_p3, p3);
+        _mm_storeu_si128((__m128i *) (s - 4 * p), p3);
+
+        p2 = _mm_andnot_si128(flat2, p2);
+        flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+        p2 = _mm_or_si128(flat2_p2, p2);
+        _mm_storeu_si128((__m128i *) (s - 3 * p), p2);
+
+        p1 = _mm_andnot_si128(flat2, p1);
+        flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+        p1 = _mm_or_si128(flat2_p1, p1);
+        _mm_storeu_si128((__m128i *) (s - 2 * p), p1);
+
+        p0 = _mm_andnot_si128(flat2, p0);
+        flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+        p0 = _mm_or_si128(flat2_p0, p0);
+        _mm_storeu_si128((__m128i *) (s - 1 * p), p0);
+
+        q0 = _mm_andnot_si128(flat2, q0);
+        flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+        q0 = _mm_or_si128(flat2_q0, q0);
+        _mm_storeu_si128((__m128i *) (s - 0 * p), q0);
+
+        q1 = _mm_andnot_si128(flat2, q1);
+        flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+        q1 = _mm_or_si128(flat2_q1, q1);
+        _mm_storeu_si128((__m128i *) (s + 1 * p), q1);
+
+        q2 = _mm_andnot_si128(flat2, q2);
+        flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+        q2 = _mm_or_si128(flat2_q2, q2);
+        _mm_storeu_si128((__m128i *) (s + 2 * p), q2);
+
+        q3 = _mm_andnot_si128(flat2, q3);
+        flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+        q3 = _mm_or_si128(flat2_q3, q3);
+        _mm_storeu_si128((__m128i *) (s + 3 * p), q3);
+
+        q4 = _mm_andnot_si128(flat2, q4);
+        flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+        q4 = _mm_or_si128(flat2_q4, q4);
+        _mm_storeu_si128((__m128i *) (s + 4 * p), q4);
+
+        q5 = _mm_andnot_si128(flat2, q5);
+        flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+        q5 = _mm_or_si128(flat2_q5, q5);
+        _mm_storeu_si128((__m128i *) (s + 5 * p), q5);
+
+        q6 = _mm_andnot_si128(flat2, q6);
+        flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+        q6 = _mm_or_si128(flat2_q6, q6);
+        _mm_storeu_si128((__m128i *) (s + 6 * p), q6);
+    }
+}
+
+void vp9_lpf_horizontal_16_avx2(unsigned char *s, int p,
+        const unsigned char *_blimit, const unsigned char *_limit,
+        const unsigned char *_thresh, int count) {
+    if (count == 1)
+        mb_lpf_horizontal_edge_w_avx2_8(s, p, _blimit, _limit, _thresh);
+    else
+        mb_lpf_horizontal_edge_w_avx2_16(s, p, _blimit, _limit, _thresh);
+}
diff --git a/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_sse2.c b/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
new file mode 100644
index 000000000..e321dbebe
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
@@ -0,0 +1,1580 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {
+  return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
+static void mb_lpf_horizontal_edge_w_sse2_8(unsigned char *s,
+                                            int p,
+                                            const unsigned char *_blimit,
+                                            const unsigned char *_limit,
+                                            const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+  __m128i abs_p1p0;
+
+  q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+  q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4),
+                                       (__m64 *)(s + 4 * p)));
+  q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3),
+                                       (__m64 *)(s + 3 * p)));
+  q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2),
+                                       (__m64 *)(s + 2 * p)));
+  q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1),
+                                       (__m64 *)(s + 1 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0),
+                                       (__m64 *)(s - 0 * p)));
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 =  _mm_srli_si128(abs_p1p0, 8);
+    fe = _mm_set1_epi8(0xfe);
+    ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+                        abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi16(0x1);
+    __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+    __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+    __m128i qs0 = _mm_xor_si128(p0q0, t80);
+    __m128i qs1 = _mm_xor_si128(p1q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+    __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+    filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, qs0ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vp9_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 0xB);
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 0xB);
+
+    // Filter1 >> 3
+    filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+    qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi16(filter1, t1);
+    filt = _mm_srai_epi16(filt, 1);
+    filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+                            filt);
+    filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+    qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+      flat = _mm_max_epu8(abs_p1p0, flat);
+      flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+      flat = _mm_subs_epu8(flat, one);
+      flat = _mm_cmpeq_epi8(flat, zero);
+      flat = _mm_and_si128(flat, mask);
+
+      q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+      q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5),
+                                           (__m64 *)(s + 5 * p)));
+
+      q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+      q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
+                                           (__m64 *)(s + 6 * p)));
+      flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
+
+      q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+      q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
+                                           (__m64 *)(s + 7 * p)));
+      work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i four = _mm_set1_epi16(4);
+      __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+      __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+      __m128i pixelFilter_p, pixelFilter_q;
+      __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+      __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+      p7_16 = _mm_unpacklo_epi8(q7p7, zero);;
+      p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+      p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+      p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+      p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+      p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+      p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+      p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+      q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+      q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+      q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+      q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+      q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+      q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+      q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+      q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+      pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                                    _mm_add_epi16(p4_16, p3_16));
+      pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                                    _mm_add_epi16(q4_16, q3_16));
+
+      pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+      pixelFilter_p =  _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+      pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+      pixelFilter_q =  _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+      pixelFilter_p =  _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+                                                         pixelFilter_q));
+      pixetFilter_p2p1p0 =   _mm_add_epi16(four,
+                                           _mm_add_epi16(pixetFilter_p2p1p0,
+                                                         pixetFilter_q2q1q0));
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                           _mm_add_epi16(p7_16, p0_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                                           _mm_add_epi16(q7_16, q0_16)), 4);
+      flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(p3_16, p0_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(q3_16, q0_16)), 3);
+
+      flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(p7_16, p7_16);
+      sum_q7 = _mm_add_epi16(q7_16, q7_16);
+      sum_p3 = _mm_add_epi16(p3_16, p3_16);
+      sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p1_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q1_16)), 4);
+      flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                             _mm_add_epi16(sum_p3, p1_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                             _mm_add_epi16(sum_q3, q1_16)), 3);
+      flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+      sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p2_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q2_16)), 4);
+      flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+                                           _mm_add_epi16(sum_p3, p2_16)), 3);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+                                           _mm_add_epi16(sum_q3, q2_16)), 3);
+      flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p3_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q3_16)), 4);
+      flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p4_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q4_16)), 4);
+      flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p5_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q5_16)), 4);
+      flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+      res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+                             _mm_add_epi16(sum_p7, p6_16)), 4);
+      res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+                             _mm_add_epi16(sum_q7, q6_16)), 4);
+      flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    flat = _mm_shuffle_epi32(flat, 68);
+    flat2 = _mm_shuffle_epi32(flat2, 68);
+
+    q2p2 = _mm_andnot_si128(flat, q2p2);
+    flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+    q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+    q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+    q6p6 = _mm_andnot_si128(flat2, q6p6);
+    flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+    q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+    _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+    _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
+
+    q5p5 = _mm_andnot_si128(flat2, q5p5);
+    flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+    q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+    _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+    _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
+
+    q4p4 = _mm_andnot_si128(flat2, q4p4);
+    flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+    q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+    _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+    _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
+
+    q3p3 = _mm_andnot_si128(flat2, q3p3);
+    flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+    q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+    _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+    _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
+
+    q2p2 = _mm_andnot_si128(flat2, q2p2);
+    flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+    _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+    _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
+
+    q1p1 = _mm_andnot_si128(flat2, q1p1);
+    flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+    q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+    _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
+
+    q0p0 = _mm_andnot_si128(flat2, q0p0);
+    flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+    q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+    _mm_storeh_pi((__m64 *)(s - 0 * p),  _mm_castsi128_ps(q0p0));
+  }
+}
+
+static INLINE __m128i filter_add2_sub2(const __m128i *const total,
+                                       const __m128i *const a1,
+                                       const __m128i *const a2,
+                                       const __m128i *const s1,
+                                       const __m128i *const s2) {
+  __m128i x = _mm_add_epi16(*a1, *total);
+  x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
+  return x;
+}
+
+static INLINE __m128i filter8_mask(const __m128i *const flat,
+                                   const __m128i *const other_filt,
+                                   const __m128i *const f8_lo,
+                                   const __m128i *const f8_hi) {
+  const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3),
+                                      _mm_srli_epi16(*f8_hi, 3));
+  const __m128i result = _mm_and_si128(*flat, f8);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(const __m128i *const flat,
+                                    const __m128i *const other_filt,
+                                    const __m128i *const f_lo,
+                                    const __m128i *const f_hi) {
+  const __m128i f = _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4),
+                                     _mm_srli_epi16(*f_hi, 4));
+  const __m128i result = _mm_and_si128(*flat, f);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static void mb_lpf_horizontal_edge_w_sse2_16(unsigned char *s,
+                                             int p,
+                                             const unsigned char *_blimit,
+                                             const unsigned char *_limit,
+                                             const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i p7, p6, p5;
+  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i q5, q6, q7;
+
+  __m128i op2, op1, op0, oq0, oq1, oq2;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+  p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+  p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
+  p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
+  q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+  q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+  q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vp9_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+    // loopfilter done
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter8
+    {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+                            _mm_add_epi16(p3_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+                            _mm_add_epi16(p2_lo, p1_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+                            _mm_add_epi16(p3_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+                            _mm_add_epi16(p2_hi, p1_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+      oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
+      const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+      const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+      const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+      const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+      const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+      const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+      const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
+
+      const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
+      const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+      const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+      const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+      const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+      const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+      const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
+
+      __m128i f_lo;
+      __m128i f_hi;
+
+      f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo);  // p7 * 7
+      f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1),
+                           _mm_add_epi16(p4_lo, f_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+                           _mm_add_epi16(p2_lo, p1_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+      f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
+
+      f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi);  // p7 * 7
+      f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1),
+                           _mm_add_epi16(p4_hi, f_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+                           _mm_add_epi16(p2_hi, p1_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+      f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
+
+      p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
+
+      f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi);
+      p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi);
+      p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi);
+      p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi);
+      op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi);
+      op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi);
+      op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi);
+      oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi);
+      oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi);
+      oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi);
+      q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi);
+      q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi);
+      q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi);
+      q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  }
+}
+
+// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
+void vp9_lpf_horizontal_16_sse2(unsigned char *s, int p,
+                                const unsigned char *_blimit,
+                                const unsigned char *_limit,
+                                const unsigned char *_thresh, int count) {
+  if (count == 1)
+    mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh);
+  else
+    mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh);
+}
+
+void vp9_lpf_horizontal_8_sse2(unsigned char *s, int p,
+                               const unsigned char *_blimit,
+                               const unsigned char *_limit,
+                               const unsigned char *_thresh, int count) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
+
+  (void)count;
+
+  q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+  q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    // filter_mask and hev_mask
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 =  _mm_srli_si128(abs_p1p0, 8);
+
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+                        abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0),
+                        abs_diff(q3p3, q0p0));
+    flat = _mm_max_epu8(abs_p1p0, flat);
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+    }
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i ps1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vp9_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 11);
+    filter1 = _mm_packs_epi16(filter1, filter1);
+
+    // Filter2 >> 3
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 11);
+    filter2 = _mm_packs_epi16(filter2, zero);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    filt = _mm_unpacklo_epi8(zero, filt);
+    filt = _mm_srai_epi16(filt, 9);
+    filt = _mm_packs_epi16(filt, zero);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_loadl_epi64((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_loadl_epi64((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_loadl_epi64((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
+    _mm_storel_epi64((__m128i *)(s + 0 * p), q0);
+    _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
+    _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void vp9_lpf_horizontal_8_dual_sse2(uint8_t *s, int p,
+                                    const uint8_t *_blimit0,
+                                    const uint8_t *_limit0,
+                                    const uint8_t *_thresh0,
+                                    const uint8_t *_blimit1,
+                                    const uint8_t *_limit1,
+                                    const uint8_t *_thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  {
+    const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                                          _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                                          _mm_subs_epu8(q0, q1));
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                                    _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                                    _mm_subs_epu8(q1, p1));
+    __m128i work;
+
+    // filter_mask and hev_mask
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+                                     _mm_subs_epu8(p1, p2)),
+                         _mm_or_si128(_mm_subs_epu8(p3, p2),
+                                      _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+                                     _mm_subs_epu8(q1, q2)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q2),
+                                      _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
+                                     _mm_subs_epu8(p0, p2)),
+                         _mm_or_si128(_mm_subs_epu8(q2, q0),
+                                      _mm_subs_epu8(q0, q2)));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
+                                     _mm_subs_epu8(p0, p3)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q0),
+                                      _mm_subs_epu8(q0, q3)));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    int i = 0;
+
+    do {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      src += 8;
+    } while (++i < 2);
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vp9_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_load_si128((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_load_si128((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_load_si128((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_load_si128((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_load_si128((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_load_si128((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+    _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void vp9_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0,
+                                    const unsigned char *_blimit1,
+                                    const unsigned char *_limit1,
+                                    const unsigned char *_thresh1) {
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i mask, hev, flat;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+
+  // filter_mask and hev_mask
+  {
+    const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+                                          _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+                                          _mm_subs_epu8(q0, q1));
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+                                    _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+                                    _mm_subs_epu8(q1, p1));
+    __m128i work;
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+                                     _mm_subs_epu8(p1, p2)),
+                         _mm_or_si128(_mm_subs_epu8(p3, p2),
+                                      _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+                                     _mm_subs_epu8(q1, q2)),
+                         _mm_or_si128(_mm_subs_epu8(q3, q2),
+                                      _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+                                      t80);
+    const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+                                      t80);
+    const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+                                      t80);
+    const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+                                      t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (vp9_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+  }
+}
+
+static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1,
+                                 int in_p, unsigned char *out, int out_p) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i x8, x9, x10, x11, x12, x13, x14, x15;
+
+  // Read in 16 lines
+  x0 = _mm_loadl_epi64((__m128i *)in0);
+  x8 = _mm_loadl_epi64((__m128i *)in1);
+  x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p));
+  x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p));
+  x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p));
+  x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p));
+  x3 = _mm_loadl_epi64((__m128i *)(in0 + 3*in_p));
+  x11 = _mm_loadl_epi64((__m128i *)(in1 + 3*in_p));
+  x4 = _mm_loadl_epi64((__m128i *)(in0 + 4*in_p));
+  x12 = _mm_loadl_epi64((__m128i *)(in1 + 4*in_p));
+  x5 = _mm_loadl_epi64((__m128i *)(in0 + 5*in_p));
+  x13 = _mm_loadl_epi64((__m128i *)(in1 + 5*in_p));
+  x6 = _mm_loadl_epi64((__m128i *)(in0 + 6*in_p));
+  x14 = _mm_loadl_epi64((__m128i *)(in1 + 6*in_p));
+  x7 = _mm_loadl_epi64((__m128i *)(in0 + 7*in_p));
+  x15 = _mm_loadl_epi64((__m128i *)(in1 + 7*in_p));
+
+  x0 = _mm_unpacklo_epi8(x0, x1);
+  x1 = _mm_unpacklo_epi8(x2, x3);
+  x2 = _mm_unpacklo_epi8(x4, x5);
+  x3 = _mm_unpacklo_epi8(x6, x7);
+
+  x8 = _mm_unpacklo_epi8(x8, x9);
+  x9 = _mm_unpacklo_epi8(x10, x11);
+  x10 = _mm_unpacklo_epi8(x12, x13);
+  x11 = _mm_unpacklo_epi8(x14, x15);
+
+  x4 = _mm_unpacklo_epi16(x0, x1);
+  x5 = _mm_unpacklo_epi16(x2, x3);
+  x12 = _mm_unpacklo_epi16(x8, x9);
+  x13 = _mm_unpacklo_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store first 4-line result
+  _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
+
+  x4 = _mm_unpackhi_epi16(x0, x1);
+  x5 = _mm_unpackhi_epi16(x2, x3);
+  x12 = _mm_unpackhi_epi16(x8, x9);
+  x13 = _mm_unpackhi_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store second 4-line result
+  _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
+}
+
+static INLINE void transpose(unsigned char *src[], int in_p,
+                             unsigned char *dst[], int out_p,
+                             int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    unsigned char *in = src[idx8x8];
+    unsigned char *out = dst[idx8x8];
+
+    x0 = _mm_loadl_epi64((__m128i *)(in + 0*in_p));  // 00 01 02 03 04 05 06 07
+    x1 = _mm_loadl_epi64((__m128i *)(in + 1*in_p));  // 10 11 12 13 14 15 16 17
+    x2 = _mm_loadl_epi64((__m128i *)(in + 2*in_p));  // 20 21 22 23 24 25 26 27
+    x3 = _mm_loadl_epi64((__m128i *)(in + 3*in_p));  // 30 31 32 33 34 35 36 37
+    x4 = _mm_loadl_epi64((__m128i *)(in + 4*in_p));  // 40 41 42 43 44 45 46 47
+    x5 = _mm_loadl_epi64((__m128i *)(in + 5*in_p));  // 50 51 52 53 54 55 56 57
+    x6 = _mm_loadl_epi64((__m128i *)(in + 6*in_p));  // 60 61 62 63 64 65 66 67
+    x7 = _mm_loadl_epi64((__m128i *)(in + 7*in_p));  // 70 71 72 73 74 75 76 77
+    // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+    x0 = _mm_unpacklo_epi8(x0, x1);
+    // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+    x1 = _mm_unpacklo_epi8(x2, x3);
+    // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+    x2 = _mm_unpacklo_epi8(x4, x5);
+    // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+    x3 = _mm_unpacklo_epi8(x6, x7);
+    // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    x4 = _mm_unpacklo_epi16(x0, x1);
+    // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+    x5 = _mm_unpacklo_epi16(x2, x3);
+    // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi32(x4, x5);
+
+    _mm_storel_pd((double *)(out + 0*out_p),
+                  _mm_castsi128_pd(x6));  // 00 10 20 30 40 50 60 70
+    _mm_storeh_pd((double *)(out + 1*out_p),
+                  _mm_castsi128_pd(x6));  // 01 11 21 31 41 51 61 71
+    _mm_storel_pd((double *)(out + 2*out_p),
+                  _mm_castsi128_pd(x7));  // 02 12 22 32 42 52 62 72
+    _mm_storeh_pd((double *)(out + 3*out_p),
+                  _mm_castsi128_pd(x7));  // 03 13 23 33 43 53 63 73
+
+    // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi16(x0, x1);
+    // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi16(x2, x3);
+    // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi32(x4, x5);
+
+    _mm_storel_pd((double *)(out + 4*out_p),
+                  _mm_castsi128_pd(x6));  // 04 14 24 34 44 54 64 74
+    _mm_storeh_pd((double *)(out + 5*out_p),
+                  _mm_castsi128_pd(x6));  // 05 15 25 35 45 55 65 75
+    _mm_storel_pd((double *)(out + 6*out_p),
+                  _mm_castsi128_pd(x7));  // 06 16 26 36 46 56 66 76
+    _mm_storeh_pd((double *)(out + 7*out_p),
+                  _mm_castsi128_pd(x7));  // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+void vp9_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vp9_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+}
+
+void vp9_lpf_vertical_8_sse2(unsigned char *s, int p,
+                             const unsigned char *blimit,
+                             const unsigned char *limit,
+                             const unsigned char *thresh, int count) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]);
+  unsigned char *src[1];
+  unsigned char *dst[1];
+  (void)count;
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  vp9_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 1);
+}
+
+void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0,
+                                  const uint8_t *thresh0,
+                                  const uint8_t *blimit1,
+                                  const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  vp9_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+}
+
+void vp9_lpf_vertical_16_sse2(unsigned char *s, int p,
+                              const unsigned char *blimit,
+                              const unsigned char *limit,
+                              const unsigned char *thresh) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit, thresh);
+
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 2);
+}
+
+void vp9_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+  // Transpose 16x16
+  transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  // Loop filtering
+  mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
+                                   thresh);
+
+  // Transpose back
+  transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/media/libvpx/vp9/common/x86/vp9_loopfilter_mmx.asm b/media/libvpx/vp9/common/x86/vp9_loopfilter_mmx.asm
new file mode 100644
index 000000000..f5f7d5af7
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_loopfilter_mmx.asm
@@ -0,0 +1,611 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void vp9_lpf_horizontal_4_mmx
+;(
+;    unsigned char *src_ptr,
+;    int src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int  count
+;)
+global sym(vp9_lpf_horizontal_4_mmx) PRIVATE
+sym(vp9_lpf_horizontal_4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 32                         ; reserve 32 bytes
+    %define t0 [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1 [rsp + 16]   ;__declspec(align(16)) char t1[8];
+
+        mov         rsi, arg(0) ;src_ptr
+        movsxd      rax, dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        movsxd      rcx, dword ptr arg(5) ;count
+.next8_h:
+        mov         rdx, arg(3) ;limit
+        movq        mm7, [rdx]
+        mov         rdi, rsi              ; rdi points to row +1 for indirect addressing
+        add         rdi, rax
+
+        ; calculate breakout conditions
+        movq        mm2, [rdi+2*rax]      ; q3
+        movq        mm1, [rsi+2*rax]      ; q2
+        movq        mm6, mm1              ; q2
+        psubusb     mm1, mm2              ; q2-=q3
+        psubusb     mm2, mm6              ; q3-=q2
+        por         mm1, mm2              ; abs(q3-q2)
+        psubusb     mm1, mm7              ;
+
+
+        movq        mm4, [rsi+rax]        ; q1
+        movq        mm3, mm4              ; q1
+        psubusb     mm4, mm6              ; q1-=q2
+        psubusb     mm6, mm3              ; q2-=q1
+        por         mm4, mm6              ; abs(q2-q1)
+
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm4, [rsi]            ; q0
+        movq        mm0, mm4              ; q0
+        psubusb     mm4, mm3              ; q0-=q1
+        psubusb     mm3, mm0              ; q1-=q0
+        por         mm4, mm3              ; abs(q0-q1)
+        movq        t0, mm4               ; save to t0
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        neg         rax                   ; negate pitch to deal with above border
+
+        movq        mm2, [rsi+4*rax]      ; p3
+        movq        mm4, [rdi+4*rax]      ; p2
+        movq        mm5, mm4              ; p2
+        psubusb     mm4, mm2              ; p2-=p3
+        psubusb     mm2, mm5              ; p3-=p2
+        por         mm4, mm2              ; abs(p3 - p2)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+
+        movq        mm4, [rsi+2*rax]      ; p1
+        movq        mm3, mm4              ; p1
+        psubusb     mm4, mm5              ; p1-=p2
+        psubusb     mm5, mm3              ; p2-=p1
+        por         mm4, mm5              ; abs(p2 - p1)
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm2, mm3              ; p1
+
+        movq        mm4, [rsi+rax]        ; p0
+        movq        mm5, mm4              ; p0
+        psubusb     mm4, mm3              ; p0-=p1
+        psubusb     mm3, mm5              ; p1-=p0
+        por         mm4, mm3              ; abs(p1 - p0)
+        movq        t1, mm4               ; save to t1
+        psubusb     mm4, mm7
+        por        mm1, mm4
+
+        movq        mm3, [rdi]            ; q1
+        movq        mm4, mm3              ; q1
+        psubusb     mm3, mm2              ; q1-=p1
+        psubusb     mm2, mm4              ; p1-=q1
+        por         mm2, mm3              ; abs(p1-q1)
+        pand        mm2, [GLOBAL(tfe)]    ; set lsb of each byte to zero
+        psrlw       mm2, 1                ; abs(p1-q1)/2
+
+        movq        mm6, mm5              ; p0
+        movq        mm3, [rsi]            ; q0
+        psubusb     mm5, mm3              ; p0-=q0
+        psubusb     mm3, mm6              ; q0-=p0
+        por         mm5, mm3              ; abs(p0 - q0)
+        paddusb     mm5, mm5              ; abs(p0-q0)*2
+        paddusb     mm5, mm2              ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        mov         rdx, arg(2) ;blimit           ; get blimit
+        movq        mm7, [rdx]            ; blimit
+
+        psubusb     mm5,    mm7           ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,    mm5
+        pxor        mm5,    mm5
+        pcmpeqb     mm1,    mm5           ; mask mm1
+
+        ; calculate high edge variance
+        mov         rdx, arg(4) ;thresh           ; get thresh
+        movq        mm7, [rdx]            ;
+        movq        mm4, t0               ; get abs (q1 - q0)
+        psubusb     mm4, mm7
+        movq        mm3, t1               ; get abs (p1 - p0)
+        psubusb     mm3, mm7
+        paddb       mm4, mm3              ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+        pcmpeqb     mm4,        mm5
+
+        pcmpeqb     mm5,        mm5
+        pxor        mm4,        mm5
+
+
+        ; start work on filters
+        movq        mm2, [rsi+2*rax]      ; p1
+        movq        mm7, [rdi]            ; q1
+        pxor        mm2, [GLOBAL(t80)]    ; p1 offset to convert to signed values
+        pxor        mm7, [GLOBAL(t80)]    ; q1 offset to convert to signed values
+        psubsb      mm2, mm7              ; p1 - q1
+        pand        mm2, mm4              ; high var mask (hvm)(p1 - q1)
+        pxor        mm6, [GLOBAL(t80)]    ; offset to convert to signed values
+        pxor        mm0, [GLOBAL(t80)]    ; offset to convert to signed values
+        movq        mm3, mm0              ; q0
+        psubsb      mm0, mm6              ; q0 - p0
+        paddsb      mm2, mm0              ; 1 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2, mm0              ; 2 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2, mm0              ; 3 * (q0 - p0) + hvm(p1 - q1)
+        pand        mm1, mm2                  ; mask filter values we don't care about
+        movq        mm2, mm1
+        paddsb      mm1, [GLOBAL(t4)]     ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+        paddsb      mm2, [GLOBAL(t3)]     ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+
+        pxor        mm0, mm0             ;
+        pxor        mm5, mm5
+        punpcklbw   mm0, mm2            ;
+        punpckhbw   mm5, mm2            ;
+        psraw       mm0, 11             ;
+        psraw       mm5, 11
+        packsswb    mm0, mm5
+        movq        mm2, mm0            ;  (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+        pxor        mm0, mm0              ; 0
+        movq        mm5, mm1              ; abcdefgh
+        punpcklbw   mm0, mm1              ; e0f0g0h0
+        psraw       mm0, 11               ; sign extended shift right by 3
+        pxor        mm1, mm1              ; 0
+        punpckhbw   mm1, mm5              ; a0b0c0d0
+        psraw       mm1, 11               ; sign extended shift right by 3
+        movq        mm5, mm0              ; save results
+
+        packsswb    mm0, mm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+        paddsw      mm5, [GLOBAL(ones)]
+        paddsw      mm1, [GLOBAL(ones)]
+        psraw       mm5, 1                ; partial shifted one more time for 2nd tap
+        psraw       mm1, 1                ; partial shifted one more time for 2nd tap
+        packsswb    mm5, mm1              ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+        pandn       mm4, mm5              ; high edge variance additive
+
+        paddsb      mm6, mm2              ; p0+= p0 add
+        pxor        mm6, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi+rax], mm6        ; write back
+
+        movq        mm6, [rsi+2*rax]      ; p1
+        pxor        mm6, [GLOBAL(t80)]    ; reoffset
+        paddsb      mm6, mm4              ; p1+= p1 add
+        pxor        mm6, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi+2*rax], mm6      ; write back
+
+        psubsb      mm3, mm0              ; q0-= q0 add
+        pxor        mm3, [GLOBAL(t80)]    ; unoffset
+        movq        [rsi], mm3            ; write back
+
+        psubsb      mm7, mm4              ; q1-= q1 add
+        pxor        mm7, [GLOBAL(t80)]    ; unoffset
+        movq        [rdi], mm7            ; write back
+
+        add         rsi,8
+        neg         rax
+        dec         rcx
+        jnz         .next8_h
+
+    add rsp, 32
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void vp9_lpf_vertical_4_mmx
+;(
+;    unsigned char *src_ptr,
+;    int  src_pixel_step,
+;    const char *blimit,
+;    const char *limit,
+;    const char *thresh,
+;    int count
+;)
+global sym(vp9_lpf_vertical_4_mmx) PRIVATE
+sym(vp9_lpf_vertical_4_mmx):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub          rsp, 64      ; reserve 64 bytes
+    %define t0   [rsp + 0]    ;__declspec(align(16)) char t0[8];
+    %define t1   [rsp + 16]   ;__declspec(align(16)) char t1[8];
+    %define srct [rsp + 32]   ;__declspec(align(16)) char srct[32];
+
+        mov         rsi,        arg(0) ;src_ptr
+        movsxd      rax,        dword ptr arg(1) ;src_pixel_step     ; destination pitch?
+
+        lea         rsi,        [rsi + rax*4 - 4]
+
+        movsxd      rcx,        dword ptr arg(5) ;count
+.next8_v:
+        mov         rdi,        rsi           ; rdi points to row +1 for indirect addressing
+        add         rdi,        rax
+
+
+        ;transpose
+        movq        mm6,        [rsi+2*rax]                 ; 67 66 65 64 63 62 61 60
+        movq        mm7,        mm6                         ; 77 76 75 74 73 72 71 70
+
+        punpckhbw   mm7,        [rdi+2*rax]                 ; 77 67 76 66 75 65 74 64
+        punpcklbw   mm6,        [rdi+2*rax]                 ; 73 63 72 62 71 61 70 60
+
+        movq        mm4,        [rsi]                       ; 47 46 45 44 43 42 41 40
+        movq        mm5,        mm4                         ; 47 46 45 44 43 42 41 40
+
+        punpckhbw   mm5,        [rsi+rax]                   ; 57 47 56 46 55 45 54 44
+        punpcklbw   mm4,        [rsi+rax]                   ; 53 43 52 42 51 41 50 40
+
+        movq        mm3,        mm5                         ; 57 47 56 46 55 45 54 44
+        punpckhwd   mm5,        mm7                         ; 77 67 57 47 76 66 56 46
+
+        punpcklwd   mm3,        mm7                         ; 75 65 55 45 74 64 54 44
+        movq        mm2,        mm4                         ; 53 43 52 42 51 41 50 40
+
+        punpckhwd   mm4,        mm6                         ; 73 63 53 43 72 62 52 42
+        punpcklwd   mm2,        mm6                         ; 71 61 51 41 70 60 50 40
+
+        neg         rax
+        movq        mm6,        [rsi+rax*2]                 ; 27 26 25 24 23 22 21 20
+
+        movq        mm1,        mm6                         ; 27 26 25 24 23 22 21 20
+        punpckhbw   mm6,        [rsi+rax]                   ; 37 27 36 36 35 25 34 24
+
+        punpcklbw   mm1,        [rsi+rax]                   ; 33 23 32 22 31 21 30 20
+        movq        mm7,        [rsi+rax*4];                ; 07 06 05 04 03 02 01 00
+
+        punpckhbw   mm7,        [rdi+rax*4]                 ; 17 07 16 06 15 05 14 04
+        movq        mm0,        mm7                         ; 17 07 16 06 15 05 14 04
+
+        punpckhwd   mm7,        mm6                         ; 37 27 17 07 36 26 16 06
+        punpcklwd   mm0,        mm6                         ; 35 25 15 05 34 24 14 04
+
+        movq        mm6,        mm7                         ; 37 27 17 07 36 26 16 06
+        punpckhdq   mm7,        mm5                         ; 77 67 57 47 37 27 17 07  = q3
+
+        punpckldq   mm6,        mm5                         ; 76 66 56 46 36 26 16 06  = q2
+
+        movq        mm5,        mm6                         ; 76 66 56 46 36 26 16 06
+        psubusb     mm5,        mm7                         ; q2-q3
+
+        psubusb     mm7,        mm6                         ; q3-q2
+        por         mm7,        mm5;                        ; mm7=abs (q3-q2)
+
+        movq        mm5,        mm0                         ; 35 25 15 05 34 24 14 04
+        punpckhdq   mm5,        mm3                         ; 75 65 55 45 35 25 15 05 = q1
+
+        punpckldq   mm0,        mm3                         ; 74 64 54 44 34 24 15 04 = q0
+        movq        mm3,        mm5                         ; 75 65 55 45 35 25 15 05 = q1
+
+        psubusb     mm3,        mm6                         ; q1-q2
+        psubusb     mm6,        mm5                         ; q2-q1
+
+        por         mm6,        mm3                         ; mm6=abs(q2-q1)
+        lea         rdx,        srct
+
+        movq        [rdx+24],   mm5                         ; save q1
+        movq        [rdx+16],   mm0                         ; save q0
+
+        movq        mm3,        [rsi+rax*4]                 ; 07 06 05 04 03 02 01 00
+        punpcklbw   mm3,        [rdi+rax*4]                 ; 13 03 12 02 11 01 10 00
+
+        movq        mm0,        mm3                         ; 13 03 12 02 11 01 10 00
+        punpcklwd   mm0,        mm1                         ; 31 21 11 01 30 20 10 00
+
+        punpckhwd   mm3,        mm1                         ; 33 23 13 03 32 22 12 02
+        movq        mm1,        mm0                         ; 31 21 11 01 30 20 10 00
+
+        punpckldq   mm0,        mm2                         ; 70 60 50 40 30 20 10 00  =p3
+        punpckhdq   mm1,        mm2                         ; 71 61 51 41 31 21 11 01  =p2
+
+        movq        mm2,        mm1                         ; 71 61 51 41 31 21 11 01  =p2
+        psubusb     mm2,        mm0                         ; p2-p3
+
+        psubusb     mm0,        mm1                         ; p3-p2
+        por         mm0,        mm2                         ; mm0=abs(p3-p2)
+
+        movq        mm2,        mm3                         ; 33 23 13 03 32 22 12 02
+        punpckldq   mm2,        mm4                         ; 72 62 52 42 32 22 12 02 = p1
+
+        punpckhdq   mm3,        mm4                         ; 73 63 53 43 33 23 13 03 = p0
+        movq        [rdx+8],    mm3                         ; save p0
+
+        movq        [rdx],      mm2                         ; save p1
+        movq        mm5,        mm2                         ; mm5 = p1
+
+        psubusb     mm2,        mm1                         ; p1-p2
+        psubusb     mm1,        mm5                         ; p2-p1
+
+        por         mm1,        mm2                         ; mm1=abs(p2-p1)
+        mov         rdx,        arg(3) ;limit
+
+        movq        mm4,        [rdx]                       ; mm4 = limit
+        psubusb     mm7,        mm4
+
+        psubusb     mm0,        mm4
+        psubusb     mm1,        mm4
+
+        psubusb     mm6,        mm4
+        por         mm7,        mm6
+
+        por         mm0,        mm1
+        por         mm0,        mm7                         ;   abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
+
+        movq        mm1,        mm5                         ; p1
+
+        movq        mm7,        mm3                         ; mm3=mm7=p0
+        psubusb     mm7,        mm5                         ; p0 - p1
+
+        psubusb     mm5,        mm3                         ; p1 - p0
+        por         mm5,        mm7                         ; abs(p1-p0)
+
+        movq        t0,         mm5                         ; save abs(p1-p0)
+        lea         rdx,        srct
+
+        psubusb     mm5,        mm4
+        por         mm0,        mm5                         ; mm0=mask
+
+        movq        mm5,        [rdx+16]                    ; mm5=q0
+        movq        mm7,        [rdx+24]                    ; mm7=q1
+
+        movq        mm6,        mm5                         ; mm6=q0
+        movq        mm2,        mm7                         ; q1
+        psubusb     mm5,        mm7                         ; q0-q1
+
+        psubusb     mm7,        mm6                         ; q1-q0
+        por         mm7,        mm5                         ; abs(q1-q0)
+
+        movq        t1,         mm7                         ; save abs(q1-q0)
+        psubusb     mm7,        mm4
+
+        por         mm0,        mm7                         ; mask
+
+        movq        mm5,        mm2                         ; q1
+        psubusb     mm5,        mm1                         ; q1-=p1
+        psubusb     mm1,        mm2                         ; p1-=q1
+        por         mm5,        mm1                         ; abs(p1-q1)
+        pand        mm5,        [GLOBAL(tfe)]               ; set lsb of each byte to zero
+        psrlw       mm5,        1                           ; abs(p1-q1)/2
+
+        mov         rdx,        arg(2) ;blimit                      ;
+
+        movq        mm4,        [rdx]                       ;blimit
+        movq        mm1,        mm3                         ; mm1=mm3=p0
+
+        movq        mm7,        mm6                         ; mm7=mm6=q0
+        psubusb     mm1,        mm7                         ; p0-q0
+
+        psubusb     mm7,        mm3                         ; q0-p0
+        por         mm1,        mm7                         ; abs(q0-p0)
+        paddusb     mm1,        mm1                         ; abs(q0-p0)*2
+        paddusb     mm1,        mm5                         ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+        psubusb     mm1,        mm4                         ; abs (p0 - q0) *2 + abs(p1-q1)/2  > blimit
+        por         mm1,        mm0;                        ; mask
+
+        pxor        mm0,        mm0
+        pcmpeqb     mm1,        mm0
+
+        ; calculate high edge variance
+        mov         rdx,        arg(4) ;thresh            ; get thresh
+        movq        mm7,        [rdx]
+        ;
+        movq        mm4,        t0              ; get abs (q1 - q0)
+        psubusb     mm4,        mm7
+
+        movq        mm3,        t1              ; get abs (p1 - p0)
+        psubusb     mm3,        mm7
+
+        por         mm4,        mm3             ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+        pcmpeqb     mm4,        mm0
+
+        pcmpeqb     mm0,        mm0
+        pxor        mm4,        mm0
+
+
+
+        ; start work on filters
+        lea         rdx,        srct
+
+        movq        mm2,        [rdx]           ; p1
+        movq        mm7,        [rdx+24]        ; q1
+
+        movq        mm6,        [rdx+8]         ; p0
+        movq        mm0,        [rdx+16]        ; q0
+
+        pxor        mm2,        [GLOBAL(t80)]   ; p1 offset to convert to signed values
+        pxor        mm7,        [GLOBAL(t80)]   ; q1 offset to convert to signed values
+
+        psubsb      mm2,        mm7             ; p1 - q1
+        pand        mm2,        mm4             ; high var mask (hvm)(p1 - q1)
+
+        pxor        mm6,        [GLOBAL(t80)]   ; offset to convert to signed values
+        pxor        mm0,        [GLOBAL(t80)]   ; offset to convert to signed values
+
+        movq        mm3,        mm0             ; q0
+        psubsb      mm0,        mm6             ; q0 - p0
+
+        paddsb      mm2,        mm0             ; 1 * (q0 - p0) + hvm(p1 - q1)
+        paddsb      mm2,        mm0             ; 2 * (q0 - p0) + hvm(p1 - q1)
+
+        paddsb      mm2,        mm0             ; 3 * (q0 - p0) + hvm(p1 - q1)
+        pand       mm1,        mm2              ; mask filter values we don't care about
+
+        movq        mm2,        mm1
+        paddsb      mm1,        [GLOBAL(t4)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+
+        paddsb      mm2,        [GLOBAL(t3)]      ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+        pxor        mm0,        mm0          ;
+
+        pxor        mm5,        mm5
+        punpcklbw   mm0,        mm2         ;
+
+        punpckhbw   mm5,        mm2         ;
+        psraw       mm0,        11              ;
+
+        psraw       mm5,        11
+        packsswb    mm0,        mm5
+
+        movq        mm2,        mm0         ;  (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+        pxor        mm0,        mm0           ; 0
+        movq        mm5,        mm1           ; abcdefgh
+
+        punpcklbw   mm0,        mm1           ; e0f0g0h0
+        psraw       mm0,        11                ; sign extended shift right by 3
+
+        pxor        mm1,        mm1           ; 0
+        punpckhbw   mm1,        mm5           ; a0b0c0d0
+
+        psraw       mm1,        11                ; sign extended shift right by 3
+        movq        mm5,        mm0              ; save results
+
+        packsswb    mm0,        mm1           ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+        paddsw      mm5,        [GLOBAL(ones)]
+
+        paddsw      mm1,        [GLOBAL(ones)]
+        psraw       mm5,        1                 ; partial shifted one more time for 2nd tap
+
+        psraw       mm1,        1                 ; partial shifted one more time for 2nd tap
+        packsswb    mm5,        mm1           ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+
+        pandn       mm4,        mm5             ; high edge variance additive
+
+        paddsb      mm6,        mm2             ; p0+= p0 add
+        pxor        mm6,        [GLOBAL(t80)]   ; unoffset
+
+        ; mm6=p0                               ;
+        movq        mm1,        [rdx]           ; p1
+        pxor        mm1,        [GLOBAL(t80)]   ; reoffset
+
+        paddsb      mm1,        mm4                 ; p1+= p1 add
+        pxor        mm1,        [GLOBAL(t80)]       ; unoffset
+        ; mm6 = p0 mm1 = p1
+
+        psubsb      mm3,        mm0                 ; q0-= q0 add
+        pxor        mm3,        [GLOBAL(t80)]       ; unoffset
+
+        ; mm3 = q0
+        psubsb      mm7,        mm4                 ; q1-= q1 add
+        pxor        mm7,        [GLOBAL(t80)]       ; unoffset
+        ; mm7 = q1
+
+        ; transpose and write back
+        ; mm1 =    72 62 52 42 32 22 12 02
+        ; mm6 =    73 63 53 43 33 23 13 03
+        ; mm3 =    74 64 54 44 34 24 14 04
+        ; mm7 =    75 65 55 45 35 25 15 05
+
+        movq        mm2,        mm1             ; 72 62 52 42 32 22 12 02
+        punpcklbw   mm2,        mm6             ; 33 32 23 22 13 12 03 02
+
+        movq        mm4,        mm3             ; 74 64 54 44 34 24 14 04
+        punpckhbw   mm1,        mm6             ; 73 72 63 62 53 52 43 42
+
+        punpcklbw   mm4,        mm7             ; 35 34 25 24 15 14 05 04
+        punpckhbw   mm3,        mm7             ; 75 74 65 64 55 54 45 44
+
+        movq        mm6,        mm2             ; 33 32 23 22 13 12 03 02
+        punpcklwd   mm2,        mm4             ; 15 14 13 12 05 04 03 02
+
+        punpckhwd   mm6,        mm4             ; 35 34 33 32 25 24 23 22
+        movq        mm5,        mm1             ; 73 72 63 62 53 52 43 42
+
+        punpcklwd   mm1,        mm3             ; 55 54 53 52 45 44 43 42
+        punpckhwd   mm5,        mm3             ; 75 74 73 72 65 64 63 62
+
+
+        ; mm2 = 15 14 13 12 05 04 03 02
+        ; mm6 = 35 34 33 32 25 24 23 22
+        ; mm5 = 55 54 53 52 45 44 43 42
+        ; mm1 = 75 74 73 72 65 64 63 62
+
+
+
+        movd        [rsi+rax*4+2], mm2
+        psrlq       mm2,        32
+
+        movd        [rdi+rax*4+2], mm2
+        movd        [rsi+rax*2+2], mm6
+
+        psrlq       mm6,        32
+        movd        [rsi+rax+2],mm6
+
+        movd        [rsi+2],    mm1
+        psrlq       mm1,        32
+
+        movd        [rdi+2],    mm1
+        neg         rax
+
+        movd        [rdi+rax+2],mm5
+        psrlq       mm5,        32
+
+        movd        [rdi+rax*2+2], mm5
+
+        lea         rsi,        [rsi+rax*8]
+        dec         rcx
+        jnz         .next8_v
+
+    add rsp, 64
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+tfe:
+    times 8 db 0xfe
+align 16
+t80:
+    times 8 db 0x80
+align 16
+t3:
+    times 8 db 0x03
+align 16
+t4:
+    times 8 db 0x04
+align 16
+ones:
+    times 4 dw 0x0001
diff --git a/media/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm b/media/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm
new file mode 100644
index 000000000..6029420d1
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_mfqe_sse2.asm
@@ -0,0 +1,287 @@
+;
+;  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+;  This file is a duplicate of mfqe_sse2.asm in VP8.
+;  TODO(jackychen): Find a way to fix the duplicate.
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp9_filter_by_weight16x16_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp9_filter_by_weight16x16_sse2) PRIVATE
+sym(vp9_filter_by_weight16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 16                     ; loop count
+    pxor        xmm6, xmm6
+
+.combine
+    movdqa      xmm2, [rax]
+    movdqa      xmm4, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    movdqa      xmm3, xmm2
+    punpcklbw   xmm2, xmm6
+    punpckhbw   xmm3, xmm6
+    pmullw      xmm2, xmm0
+    pmullw      xmm3, xmm0
+
+    ; dst * dst_weight
+    movdqa      xmm5, xmm4
+    punpcklbw   xmm4, xmm6
+    punpckhbw   xmm5, xmm6
+    pmullw      xmm4, xmm1
+    pmullw      xmm5, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm4
+    paddw       xmm3, xmm5
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    paddw       xmm3, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+    psrlw       xmm3, 4
+
+    packuswb    xmm2, xmm3
+    movdqa      [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp9_filter_by_weight8x8_sse2
+;(
+;    unsigned char *src,
+;    int            src_stride,
+;    unsigned char *dst,
+;    int            dst_stride,
+;    int            src_weight
+;)
+global sym(vp9_filter_by_weight8x8_sse2) PRIVATE
+sym(vp9_filter_by_weight8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    movd        xmm0, arg(4)                ; src_weight
+    pshuflw     xmm0, xmm0, 0x0             ; replicate to all low words
+    punpcklqdq  xmm0, xmm0                  ; replicate to all hi words
+
+    movdqa      xmm1, [GLOBAL(tMFQE)]
+    psubw       xmm1, xmm0                  ; dst_weight
+
+    mov         rax, arg(0)                 ; src
+    mov         rsi, arg(1)                 ; src_stride
+    mov         rdx, arg(2)                 ; dst
+    mov         rdi, arg(3)                 ; dst_stride
+
+    mov         rcx, 8                      ; loop count
+    pxor        xmm4, xmm4
+
+.combine
+    movq        xmm2, [rax]
+    movq        xmm3, [rdx]
+    add         rax, rsi
+
+    ; src * src_weight
+    punpcklbw   xmm2, xmm4
+    pmullw      xmm2, xmm0
+
+    ; dst * dst_weight
+    punpcklbw   xmm3, xmm4
+    pmullw      xmm3, xmm1
+
+    ; sum, round and shift
+    paddw       xmm2, xmm3
+    paddw       xmm2, [GLOBAL(tMFQE_round)]
+    psrlw       xmm2, 4
+
+    packuswb    xmm2, xmm4
+    movq        [rdx], xmm2
+    add         rdx, rdi
+
+    dec         rcx
+    jnz         .combine
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+
+    ret
+
+;void vp9_variance_and_sad_16x16_sse2 | arg
+;(
+;    unsigned char *src1,          0
+;    int            stride1,       1
+;    unsigned char *src2,          2
+;    int            stride2,       3
+;    unsigned int  *variance,      4
+;    unsigned int  *sad,           5
+;)
+global sym(vp9_variance_and_sad_16x16_sse2) PRIVATE
+sym(vp9_variance_and_sad_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov         rax,        arg(0)          ; src1
+    mov         rcx,        arg(1)          ; stride1
+    mov         rdx,        arg(2)          ; src2
+    mov         rdi,        arg(3)          ; stride2
+
+    mov         rsi,        16              ; block height
+
+    ; Prep accumulator registers
+    pxor        xmm3, xmm3                  ; SAD
+    pxor        xmm4, xmm4                  ; sum of src2
+    pxor        xmm5, xmm5                  ; sum of src2^2
+
+    ; Because we're working with the actual output frames
+    ; we can't depend on any kind of data alignment.
+.accumulate
+    movdqa      xmm0, [rax]                 ; src1
+    movdqa      xmm1, [rdx]                 ; src2
+    add         rax, rcx                    ; src1 + stride1
+    add         rdx, rdi                    ; src2 + stride2
+
+    ; SAD(src1, src2)
+    psadbw      xmm0, xmm1
+    paddusw     xmm3, xmm0
+
+    ; SUM(src2)
+    pxor        xmm2, xmm2
+    psadbw      xmm2, xmm1                  ; sum src2 by misusing SAD against 0
+    paddusw     xmm4, xmm2
+
+    ; pmaddubsw would be ideal if it took two unsigned values. instead,
+    ; it expects a signed and an unsigned value. so instead we zero extend
+    ; and operate on words.
+    pxor        xmm2, xmm2
+    movdqa      xmm0, xmm1
+    punpcklbw   xmm0, xmm2
+    punpckhbw   xmm1, xmm2
+    pmaddwd     xmm0, xmm0
+    pmaddwd     xmm1, xmm1
+    paddd       xmm5, xmm0
+    paddd       xmm5, xmm1
+
+    sub         rsi,        1
+    jnz         .accumulate
+
+    ; phaddd only operates on adjacent double words.
+    ; Finalize SAD and store
+    movdqa      xmm0, xmm3
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm3
+    paddd       xmm0, [GLOBAL(t128)]
+    psrld       xmm0, 8
+
+    mov         rax,  arg(5)
+    movd        [rax], xmm0
+
+    ; Accumulate sum of src2
+    movdqa      xmm0, xmm4
+    psrldq      xmm0, 8
+    paddusw     xmm0, xmm4
+    ; Square src2. Ignore high value
+    pmuludq     xmm0, xmm0
+    psrld       xmm0, 8
+
+    ; phaddw could be used to sum adjacent values but we want
+    ; all the values summed. promote to doubles, accumulate,
+    ; shift and sum
+    pxor        xmm2, xmm2
+    movdqa      xmm1, xmm5
+    punpckldq   xmm1, xmm2
+    punpckhdq   xmm5, xmm2
+    paddd       xmm1, xmm5
+    movdqa      xmm2, xmm1
+    psrldq      xmm1, 8
+    paddd       xmm1, xmm2
+
+    psubd       xmm1, xmm0
+
+    ; (variance + 128) >> 8
+    paddd       xmm1, [GLOBAL(t128)]
+    psrld       xmm1, 8
+    mov         rax,  arg(4)
+
+    movd        [rax], xmm1
+
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+t128:
+%ifndef __NASM_VER__
+    ddq 128
+%elif CONFIG_BIG_ENDIAN
+    dq  0, 128
+%else
+    dq  128, 0
+%endif
+align 16
+tMFQE: ; 1 << MFQE_PRECISION
+    times 8 dw 0x10
+align 16
+tMFQE_round: ; 1 << (MFQE_PRECISION - 1)
+    times 8 dw 0x08
diff --git a/media/libvpx/vp9/common/x86/vp9_postproc_sse2.asm b/media/libvpx/vp9/common/x86/vp9_postproc_sse2.asm
new file mode 100644
index 000000000..ec8bfdb18
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_postproc_sse2.asm
@@ -0,0 +1,694 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;void vp9_post_proc_down_and_across_xmm
+;(
+;    unsigned char *src_ptr,
+;    unsigned char *dst_ptr,
+;    int src_pixels_per_line,
+;    int dst_pixels_per_line,
+;    int rows,
+;    int cols,
+;    int flimit
+;)
+global sym(vp9_post_proc_down_and_across_xmm) PRIVATE
+sym(vp9_post_proc_down_and_across_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+    ALIGN_STACK 16, rax
+    ; move the global rd onto the stack, since we don't have enough registers
+    ; to do PIC addressing
+    movdqa      xmm0, [GLOBAL(rd42)]
+    sub         rsp, 16
+    movdqa      [rsp], xmm0
+%define RD42 [rsp]
+%else
+%define RD42 [GLOBAL(rd42)]
+%endif
+
+
+        movd        xmm2,       dword ptr arg(6) ;flimit
+        punpcklwd   xmm2,       xmm2
+        punpckldq   xmm2,       xmm2
+        punpcklqdq  xmm2,       xmm2
+
+        mov         rsi,        arg(0) ;src_ptr
+        mov         rdi,        arg(1) ;dst_ptr
+
+        movsxd      rcx,        DWORD PTR arg(4) ;rows
+        movsxd      rax,        DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
+        pxor        xmm0,       xmm0              ; mm0 = 00000000
+
+.nextrow:
+
+        xor         rdx,        rdx       ; clear out rdx for use as loop counter
+.nextcol:
+        movq        xmm3,       QWORD PTR [rsi]         ; mm4 = r0 p0..p7
+        punpcklbw   xmm3,       xmm0                    ; mm3 = p0..p3
+        movdqa      xmm1,       xmm3                    ; mm1 = p0..p3
+        psllw       xmm3,       2                       ;
+
+        movq        xmm5,       QWORD PTR [rsi + rax]   ; mm4 = r1 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r1 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm6
+
+        ; thresholding
+        movdqa      xmm7,       xmm1                    ; mm7 = r0 p0..p3
+        psubusw     xmm7,       xmm5                    ; mm7 = r0 p0..p3 - r1 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r1 p0..p3 - r0 p0..p3
+        paddusw     xmm7,       xmm5                    ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
+        pcmpgtw     xmm7,       xmm2
+
+        movq        xmm5,       QWORD PTR [rsi + 2*rax] ; mm4 = r2 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r2 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm5                    ; mm6 = r0 p0..p3 - r2 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r2 p0..p3 - r2 p0..p3
+        paddusw     xmm6,       xmm5                    ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+
+        neg         rax
+        movq        xmm5,       QWORD PTR [rsi+2*rax]   ; mm4 = r-2 p0..p7
+        punpcklbw   xmm5,       xmm0                    ; mm5 = r-2 p0..p3
+        paddusw     xmm3,       xmm5                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm5                    ; mm6 = p0..p3 - r-2 p0..p3
+        psubusw     xmm5,       xmm1                    ; mm5 = r-2 p0..p3 - p0..p3
+        paddusw     xmm6,       xmm5                    ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+        movq        xmm4,       QWORD PTR [rsi+rax]     ; mm4 = r-1 p0..p7
+        punpcklbw   xmm4,       xmm0                    ; mm4 = r-1 p0..p3
+        paddusw     xmm3,       xmm4                    ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1                    ; mm6 = r0 p0..p3
+        psubusw     xmm6,       xmm4                    ; mm6 = p0..p3 - r-2 p0..p3
+        psubusw     xmm4,       xmm1                    ; mm5 = r-1 p0..p3 - p0..p3
+        paddusw     xmm6,       xmm4                    ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6                    ; accumulate thresholds
+
+
+        paddusw     xmm3,       RD42                    ; mm3 += round value
+        psraw       xmm3,       3                       ; mm3 /= 8
+
+        pand        xmm1,       xmm7                    ; mm1 select vals > thresh from source
+        pandn       xmm7,       xmm3                    ; mm7 select vals < thresh from blurred result
+        paddusw     xmm1,       xmm7                    ; combination
+
+        packuswb    xmm1,       xmm0                    ; pack to bytes
+        movq        QWORD PTR [rdi], xmm1             ;
+
+        neg         rax                   ; pitch is positive
+        add         rsi,        8
+        add         rdi,        8
+
+        add         rdx,        8
+        cmp         edx,        dword arg(5) ;cols
+
+        jl          .nextcol
+
+        ; done with the all cols, start the across filtering in place
+        sub         rsi,        rdx
+        sub         rdi,        rdx
+
+        xor         rdx,        rdx
+        movq        mm0,        QWORD PTR [rdi-8];
+
+.acrossnextcol:
+        movq        xmm7,       QWORD PTR [rdi +rdx -2]
+        movd        xmm4,       DWORD PTR [rdi +rdx +6]
+
+        pslldq      xmm4,       8
+        por         xmm4,       xmm7
+
+        movdqa      xmm3,       xmm4
+        psrldq      xmm3,       2
+        punpcklbw   xmm3,       xmm0              ; mm3 = p0..p3
+        movdqa      xmm1,       xmm3              ; mm1 = p0..p3
+        psllw       xmm3,       2
+
+
+        movdqa      xmm5,       xmm4
+        psrldq      xmm5,       3
+        punpcklbw   xmm5,       xmm0              ; mm5 = p1..p4
+        paddusw     xmm3,       xmm5              ; mm3 += mm6
+
+        ; thresholding
+        movdqa      xmm7,       xmm1              ; mm7 = p0..p3
+        psubusw     xmm7,       xmm5              ; mm7 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm7,       xmm5              ; mm7 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm7,       xmm2
+
+        movdqa      xmm5,       xmm4
+        psrldq      xmm5,       4
+        punpcklbw   xmm5,       xmm0              ; mm5 = p2..p5
+        paddusw     xmm3,       xmm5              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm5              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm5              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+
+        movdqa      xmm5,       xmm4              ; mm5 = p-2..p5
+        punpcklbw   xmm5,       xmm0              ; mm5 = p-2..p1
+        paddusw     xmm3,       xmm5              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm5              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm5,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm5              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+        psrldq      xmm4,       1                   ; mm4 = p-1..p5
+        punpcklbw   xmm4,       xmm0              ; mm4 = p-1..p2
+        paddusw     xmm3,       xmm4              ; mm3 += mm5
+
+        ; thresholding
+        movdqa      xmm6,       xmm1              ; mm6 = p0..p3
+        psubusw     xmm6,       xmm4              ; mm6 = p0..p3 - p1..p4
+        psubusw     xmm4,       xmm1              ; mm5 = p1..p4 - p0..p3
+        paddusw     xmm6,       xmm4              ; mm6 = abs(p0..p3 - p1..p4)
+        pcmpgtw     xmm6,       xmm2
+        por         xmm7,       xmm6              ; accumulate thresholds
+
+        paddusw     xmm3,       RD42              ; mm3 += round value
+        psraw       xmm3,       3                 ; mm3 /= 8
+
+        pand        xmm1,       xmm7              ; mm1 select vals > thresh from source
+        pandn       xmm7,       xmm3              ; mm7 select vals < thresh from blurred result
+        paddusw     xmm1,       xmm7              ; combination
+
+        packuswb    xmm1,       xmm0              ; pack to bytes
+        movq        QWORD PTR [rdi+rdx-8],  mm0   ; store previous four bytes
+        movdq2q     mm0,        xmm1
+
+        add         rdx,        8
+        cmp         edx,        dword arg(5) ;cols
+        jl          .acrossnextcol;
+
+        ; last 8 pixels
+        movq        QWORD PTR [rdi+rdx-8],  mm0
+
+        ; done with this rwo
+        add         rsi,rax               ; next line
+        mov         eax, dword arg(3) ;dst_pixels_per_line ; destination pitch?
+        add         rdi,rax               ; next destination
+        mov         eax, dword arg(2) ;src_pixels_per_line ; destination pitch?
+
+        dec         rcx                   ; decrement count
+        jnz         .nextrow              ; next row
+
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+    add rsp,16
+    pop rsp
+%endif
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef RD42
+
+
+;void vp9_mbpost_proc_down_xmm(unsigned char *dst,
+;                            int pitch, int rows, int cols,int flimit)
+extern sym(vp9_rv)
+global sym(vp9_mbpost_proc_down_xmm) PRIVATE
+sym(vp9_mbpost_proc_down_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 128+16
+
+    ; unsigned char d[16][8] at [rsp]
+    ; create flimit2 at [rsp+128]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp+128], eax
+    mov         [rsp+128+4], eax
+    mov         [rsp+128+8], eax
+    mov         [rsp+128+12], eax
+%define flimit4 [rsp+128]
+
+%if ABI_IS_32BIT=0
+    lea         r8,       [GLOBAL(sym(vp9_rv))]
+%endif
+
+    ;rows +=8;
+    add         dword arg(2), 8
+
+    ;for(c=0; c<cols; c+=8)
+.loop_col:
+            mov         rsi,        arg(0) ; s
+            pxor        xmm0,       xmm0        ;
+
+            movsxd      rax,        dword ptr arg(1) ;pitch       ;
+            neg         rax                                     ; rax = -pitch
+
+            lea         rsi,        [rsi + rax*8];              ; rdi = s[-pitch*8]
+            neg         rax
+
+
+            pxor        xmm5,       xmm5
+            pxor        xmm6,       xmm6        ;
+
+            pxor        xmm7,       xmm7        ;
+            mov         rdi,        rsi
+
+            mov         rcx,        15          ;
+
+.loop_initvar:
+            movq        xmm1,       QWORD PTR [rdi];
+            punpcklbw   xmm1,       xmm0        ;
+
+            paddw       xmm5,       xmm1        ;
+            pmullw      xmm1,       xmm1        ;
+
+            movdqa      xmm2,       xmm1        ;
+            punpcklwd   xmm1,       xmm0        ;
+
+            punpckhwd   xmm2,       xmm0        ;
+            paddd       xmm6,       xmm1        ;
+
+            paddd       xmm7,       xmm2        ;
+            lea         rdi,        [rdi+rax]   ;
+
+            dec         rcx
+            jne         .loop_initvar
+            ;save the var and sum
+            xor         rdx,        rdx
+.loop_row:
+            movq        xmm1,       QWORD PTR [rsi]     ; [s-pitch*8]
+            movq        xmm2,       QWORD PTR [rdi]     ; [s+pitch*7]
+
+            punpcklbw   xmm1,       xmm0
+            punpcklbw   xmm2,       xmm0
+
+            paddw       xmm5,       xmm2
+            psubw       xmm5,       xmm1
+
+            pmullw      xmm2,       xmm2
+            movdqa      xmm4,       xmm2
+
+            punpcklwd   xmm2,       xmm0
+            punpckhwd   xmm4,       xmm0
+
+            paddd       xmm6,       xmm2
+            paddd       xmm7,       xmm4
+
+            pmullw      xmm1,       xmm1
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm0
+            psubd       xmm6,       xmm1
+
+            punpckhwd   xmm2,       xmm0
+            psubd       xmm7,       xmm2
+
+
+            movdqa      xmm3,       xmm6
+            pslld       xmm3,       4
+
+            psubd       xmm3,       xmm6
+            movdqa      xmm1,       xmm5
+
+            movdqa      xmm4,       xmm5
+            pmullw      xmm1,       xmm1
+
+            pmulhw      xmm4,       xmm4
+            movdqa      xmm2,       xmm1
+
+            punpcklwd   xmm1,       xmm4
+            punpckhwd   xmm2,       xmm4
+
+            movdqa      xmm4,       xmm7
+            pslld       xmm4,       4
+
+            psubd       xmm4,       xmm7
+
+            psubd       xmm3,       xmm1
+            psubd       xmm4,       xmm2
+
+            psubd       xmm3,       flimit4
+            psubd       xmm4,       flimit4
+
+            psrad       xmm3,       31
+            psrad       xmm4,       31
+
+            packssdw    xmm3,       xmm4
+            packsswb    xmm3,       xmm0
+
+            movq        xmm1,       QWORD PTR [rsi+rax*8]
+
+            movq        xmm2,       xmm1
+            punpcklbw   xmm1,       xmm0
+
+            paddw       xmm1,       xmm5
+            mov         rcx,        rdx
+
+            and         rcx,        127
+%if ABI_IS_32BIT=1 && CONFIG_PIC=1
+            push        rax
+            lea         rax,        [GLOBAL(sym(vp9_rv))]
+            movdqu      xmm4,       [rax + rcx*2] ;vp9_rv[rcx*2]
+            pop         rax
+%elif ABI_IS_32BIT=0
+            movdqu      xmm4,       [r8 + rcx*2] ;vp9_rv[rcx*2]
+%else
+            movdqu      xmm4,       [sym(vp9_rv) + rcx*2]
+%endif
+
+            paddw       xmm1,       xmm4
+            ;paddw     xmm1,       eight8s
+            psraw       xmm1,       4
+
+            packuswb    xmm1,       xmm0
+            pand        xmm1,       xmm3
+
+            pandn       xmm3,       xmm2
+            por         xmm1,       xmm3
+
+            and         rcx,        15
+            movq        QWORD PTR   [rsp + rcx*8], xmm1 ;d[rcx*8]
+
+            mov         rcx,        rdx
+            sub         rcx,        8
+
+            and         rcx,        15
+            movq        mm0,        [rsp + rcx*8] ;d[rcx*8]
+
+            movq        [rsi],      mm0
+            lea         rsi,        [rsi+rax]
+
+            lea         rdi,        [rdi+rax]
+            add         rdx,        1
+
+            cmp         edx,        dword arg(2) ;rows
+            jl          .loop_row
+
+        add         dword arg(0), 8 ; s += 8
+        sub         dword arg(3), 8 ; cols -= 8
+        cmp         dword arg(3), 0
+        jg          .loop_col
+
+    add         rsp, 128+16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+;void vp9_mbpost_proc_across_ip_xmm(unsigned char *src,
+;                                int pitch, int rows, int cols,int flimit)
+global sym(vp9_mbpost_proc_across_ip_xmm) PRIVATE
+sym(vp9_mbpost_proc_across_ip_xmm):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16
+
+    ; create flimit4 at [rsp]
+    mov         eax, dword ptr arg(4) ;flimit
+    mov         [rsp], eax
+    mov         [rsp+4], eax
+    mov         [rsp+8], eax
+    mov         [rsp+12], eax
+%define flimit4 [rsp]
+
+
+    ;for(r=0;r<rows;r++)
+.ip_row_loop:
+
+        xor         rdx,    rdx ;sumsq=0;
+        xor         rcx,    rcx ;sum=0;
+        mov         rsi,    arg(0); s
+        mov         rdi,    -8
+.ip_var_loop:
+        ;for(i=-8;i<=6;i++)
+        ;{
+        ;    sumsq += s[i]*s[i];
+        ;    sum   += s[i];
+        ;}
+        movzx       eax, byte [rsi+rdi]
+        add         ecx, eax
+        mul         al
+        add         edx, eax
+        add         rdi, 1
+        cmp         rdi, 6
+        jle         .ip_var_loop
+
+
+            ;mov         rax,    sumsq
+            ;movd        xmm7,   rax
+            movd        xmm7,   edx
+
+            ;mov         rax,    sum
+            ;movd        xmm6,   rax
+            movd        xmm6,   ecx
+
+            mov         rsi,    arg(0) ;s
+            xor         rcx,    rcx
+
+            movsxd      rdx,    dword arg(3) ;cols
+            add         rdx,    8
+            pxor        mm0,    mm0
+            pxor        mm1,    mm1
+
+            pxor        xmm0,   xmm0
+.nextcol4:
+
+            movd        xmm1,   DWORD PTR [rsi+rcx-8]   ; -8 -7 -6 -5
+            movd        xmm2,   DWORD PTR [rsi+rcx+7]   ; +7 +8 +9 +10
+
+            punpcklbw   xmm1,   xmm0                    ; expanding
+            punpcklbw   xmm2,   xmm0                    ; expanding
+
+            punpcklwd   xmm1,   xmm0                    ; expanding to dwords
+            punpcklwd   xmm2,   xmm0                    ; expanding to dwords
+
+            psubd       xmm2,   xmm1                    ; 7--8   8--7   9--6 10--5
+            paddd       xmm1,   xmm1                    ; -8*2   -7*2   -6*2 -5*2
+
+            paddd       xmm1,   xmm2                    ; 7+-8   8+-7   9+-6 10+-5
+            pmaddwd     xmm1,   xmm2                    ; squared of 7+-8   8+-7   9+-6 10+-5
+
+            paddd       xmm6,   xmm2
+            paddd       xmm7,   xmm1
+
+            pshufd      xmm6,   xmm6,   0               ; duplicate the last ones
+            pshufd      xmm7,   xmm7,   0               ; duplicate the last ones
+
+            psrldq      xmm1,       4                   ; 8--7   9--6 10--5  0000
+            psrldq      xmm2,       4                   ; 8--7   9--6 10--5  0000
+
+            pshufd      xmm3,   xmm1,   3               ; 0000  8--7   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   3               ; 0000  8--7   8--7   8--7 squared
+
+            paddd       xmm6,   xmm4
+            paddd       xmm7,   xmm3
+
+            pshufd      xmm3,   xmm1,   01011111b       ; 0000  0000   9--6   9--6 squared
+            pshufd      xmm4,   xmm2,   01011111b       ; 0000  0000   9--6   9--6 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            pshufd      xmm3,   xmm1,   10111111b       ; 0000  0000   8--7   8--7 squared
+            pshufd      xmm4,   xmm2,   10111111b       ; 0000  0000   8--7   8--7 squared
+
+            paddd       xmm7,   xmm3
+            paddd       xmm6,   xmm4
+
+            movdqa      xmm3,   xmm6
+            pmaddwd     xmm3,   xmm3
+
+            movdqa      xmm5,   xmm7
+            pslld       xmm5,   4
+
+            psubd       xmm5,   xmm7
+            psubd       xmm5,   xmm3
+
+            psubd       xmm5,   flimit4
+            psrad       xmm5,   31
+
+            packssdw    xmm5,   xmm0
+            packsswb    xmm5,   xmm0
+
+            movd        xmm1,   DWORD PTR [rsi+rcx]
+            movq        xmm2,   xmm1
+
+            punpcklbw   xmm1,   xmm0
+            punpcklwd   xmm1,   xmm0
+
+            paddd       xmm1,   xmm6
+            paddd       xmm1,   [GLOBAL(four8s)]
+
+            psrad       xmm1,   4
+            packssdw    xmm1,   xmm0
+
+            packuswb    xmm1,   xmm0
+            pand        xmm1,   xmm5
+
+            pandn       xmm5,   xmm2
+            por         xmm5,   xmm1
+
+            movd        [rsi+rcx-8],  mm0
+            movq        mm0,    mm1
+
+            movdq2q     mm1,    xmm5
+            psrldq      xmm7,   12
+
+            psrldq      xmm6,   12
+            add         rcx,    4
+
+            cmp         rcx,    rdx
+            jl          .nextcol4
+
+        ;s+=pitch;
+        movsxd rax, dword arg(1)
+        add    arg(0), rax
+
+        sub dword arg(2), 1 ;rows-=1
+        cmp dword arg(2), 0
+        jg .ip_row_loop
+
+    add         rsp, 16
+    pop         rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%undef flimit4
+
+
+;void vp9_plane_add_noise_wmt (unsigned char *start, unsigned char *noise,
+;                            unsigned char blackclamp[16],
+;                            unsigned char whiteclamp[16],
+;                            unsigned char bothclamp[16],
+;                            unsigned int width, unsigned int height, int pitch)
+global sym(vp9_plane_add_noise_wmt) PRIVATE
+sym(vp9_plane_add_noise_wmt):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 8
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+.addnoise_loop:
+    call sym(LIBVPX_RAND) WRT_PLT
+    mov     rcx, arg(1) ;noise
+    and     rax, 0xff
+    add     rcx, rax
+
+    ; we rely on the fact that the clamping vectors are stored contiguously
+    ; in black/white/both order. Note that we have to reload this here because
+    ; rdx could be trashed by rand()
+    mov     rdx, arg(2) ; blackclamp
+
+
+            mov     rdi, rcx
+            movsxd  rcx, dword arg(5) ;[Width]
+            mov     rsi, arg(0) ;Pos
+            xor         rax,rax
+
+.addnoise_nextset:
+            movdqu      xmm1,[rsi+rax]         ; get the source
+
+            psubusb     xmm1, [rdx]    ;blackclamp        ; clamp both sides so we don't outrange adding noise
+            paddusb     xmm1, [rdx+32] ;bothclamp
+            psubusb     xmm1, [rdx+16] ;whiteclamp
+
+            movdqu      xmm2,[rdi+rax]         ; get the noise for this line
+            paddb       xmm1,xmm2              ; add it in
+            movdqu      [rsi+rax],xmm1         ; store the result
+
+            add         rax,16                 ; move to the next line
+
+            cmp         rax, rcx
+            jl          .addnoise_nextset
+
+    movsxd  rax, dword arg(7) ; Pitch
+    add     arg(0), rax ; Start += Pitch
+    sub     dword arg(6), 1   ; Height -= 1
+    jg      .addnoise_loop
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+SECTION_RODATA
+align 16
+rd42:
+    times 8 dw 0x04
+four8s:
+    times 4 dd 8
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
new file mode 100644
index 000000000..cee8d1e76
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_avx2.c
@@ -0,0 +1,602 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Due to a header conflict between math.h and intrinsics includes with ceil()
+// in certain configurations under vs9 this include needs to precede
+// immintrin.h.
+#include "./vp9_rtcd.h"
+
+#include <immintrin.h>
+
+#include "vp9/common/x86/convolve.h"
+#include "vpx_ports/mem.h"
+
+// filters for 16_h8 and 16_v8
+DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+#if defined(__clang__)
+# if __clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ <= 3) || \
+      (defined(__APPLE__) && __clang_major__ == 5 && __clang_minor__ == 0)
+#  define MM256_BROADCASTSI128_SI256(x) \
+       _mm_broadcastsi128_si256((__m128i const *)&(x))
+# else  // clang > 3.3, and not 5.0 on macosx.
+#  define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+# endif  // clang <= 3.3
+#elif defined(__GNUC__)
+# if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#  define MM256_BROADCASTSI128_SI256(x) \
+       _mm_broadcastsi128_si256((__m128i const *)&(x))
+# elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#  define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+# else  // gcc > 4.7
+#  define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+# endif  // gcc <= 4.6
+#else  // !(gcc || clang)
+# define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // __clang__
+
+static void vp9_filter_block1d16_h8_avx2(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1, srcReg32b2, filtersReg32;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to 8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32,
+                  _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x706u));
+
+  filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+  filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+  filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pixels_per_line << 1;
+  dst_stride = output_pitch << 1;
+  for (i = output_height; i > 1; i-=2) {
+    // load the 2 strides of source
+    srcReg32b1 = _mm256_castsi128_si256(
+                 _mm_loadu_si128((const __m128i *)(src_ptr - 3)));
+    srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+                 _mm_loadu_si128((const __m128i *)
+                 (src_ptr+src_pixels_per_line-3)), 1);
+
+    // filter the source buffer
+    srcRegFilt32b1_1= _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+                       _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // reading 2 strides of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 = _mm256_castsi128_si256(
+                 _mm_loadu_si128((const __m128i *)(src_ptr + 5)));
+    srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+                 _mm_loadu_si128((const __m128i *)
+                 (src_ptr+src_pixels_per_line+5)), 1);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1,
+                       _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // filter the source buffer
+    srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3= _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2= _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+                       _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1,
+                       _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
+
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
+    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1,
+                                           srcRegFilt32b2_1);
+
+    src_ptr+=src_stride;
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr,
+    _mm256_castsi256_si128(srcRegFilt32b1_1));
+
+    // save the next 16 bits
+    _mm_store_si128((__m128i*)(output_ptr+output_pitch),
+    _mm256_extractf128_si256(srcRegFilt32b1_1, 1));
+    output_ptr+=dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
+    __m128i srcRegFilt2, srcRegFilt3;
+
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1,
+                    _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1,
+                  _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1,
+                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg1,
+                 _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1,
+                 _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+
+    // reading the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    // filter the source buffer
+    srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2,
+                    _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1,
+                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2,
+                  _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2,
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                    _mm256_castsi256_si128(addFilterReg64));
+
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+                    _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
+    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1_1);
+  }
+}
+
+static void vp9_filter_block1d16_v8_avx2(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pitch,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t out_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64;
+  __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
+  __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
+  __m256i srcReg32b11, srcReg32b12, filtersReg32;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the
+  // same data in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32,
+                  _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32,
+                 _mm256_set1_epi16(0x706u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  // load 16 bytes 7 times in stride of src_pitch
+  srcReg32b1 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr)));
+  srcReg32b2 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
+  srcReg32b3 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
+  srcReg32b4 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
+  srcReg32b5 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
+  srcReg32b6 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
+  srcReg32b7 = _mm256_castsi128_si256(
+               _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
+
+  // have each consecutive loads on the same 256 register
+  srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+               _mm256_castsi256_si128(srcReg32b2), 1);
+  srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+               _mm256_castsi256_si128(srcReg32b3), 1);
+  srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
+               _mm256_castsi256_si128(srcReg32b4), 1);
+  srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
+               _mm256_castsi256_si128(srcReg32b5), 1);
+  srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
+               _mm256_castsi256_si128(srcReg32b6), 1);
+  srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
+               _mm256_castsi256_si128(srcReg32b7), 1);
+
+  // merge every two consecutive registers except the last one
+  srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
+  srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
+
+  // save
+  srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+
+  // save
+  srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+
+
+  for (i = output_height; i > 1; i-=2) {
+     // load the last 2 loads of 16 bytes and have every two
+     // consecutive loads in the same 256 bit register
+     srcReg32b8 = _mm256_castsi128_si256(
+     _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
+     srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+     _mm256_castsi256_si128(srcReg32b8), 1);
+     srcReg32b9 = _mm256_castsi128_si256(
+     _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8)));
+     srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
+     _mm256_castsi256_si128(srcReg32b9), 1);
+
+     // merge every two consecutive registers
+     // save
+     srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
+     srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
+     srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
+
+     // add and saturate the results together
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
+     srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
+
+     // add and saturate the results together
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                   _mm256_min_epi16(srcReg32b8, srcReg32b12));
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                   _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
+     srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
+
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
+
+     // multiply 2 adjacent elements with the filter and add the result
+     srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
+     srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
+
+     // add and saturate the results together
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                  _mm256_min_epi16(srcReg32b8, srcReg32b12));
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                  _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+     srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
+     srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
+
+     // shift by 7 bit each 16 bit
+     srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
+     srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
+
+     // shrink to 8 bit each 16 bits, the first lane contain the first
+     // convolve result and the second lane contain the second convolve
+     // result
+     srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+
+     src_ptr+=src_stride;
+
+     // save 16 bytes
+     _mm_store_si128((__m128i*)output_ptr,
+     _mm256_castsi256_si128(srcReg32b1));
+
+     // save the next 16 bits
+     _mm_store_si128((__m128i*)(output_ptr+out_pitch),
+     _mm256_extractf128_si256(srcReg32b1, 1));
+
+     output_ptr+=dst_stride;
+
+     // save part of the registers for next strides
+     srcReg32b10 = srcReg32b11;
+     srcReg32b1 = srcReg32b3;
+     srcReg32b11 = srcReg32b2;
+     srcReg32b3 = srcReg32b5;
+     srcReg32b2 = srcReg32b4;
+     srcReg32b5 = srcReg32b7;
+     srcReg32b7 = srcReg32b9;
+  }
+  if (i > 0) {
+    __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
+    __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
+    // load the last 16 bytes
+    srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the last 2 results together
+    srcRegFilt4 = _mm_unpacklo_epi8(
+                  _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+    srcRegFilt7 = _mm_unpackhi_epi8(
+                  _mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
+                  _mm256_castsi256_si128(firstFilters));
+    srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4,
+                  _mm256_castsi256_si128(forthFilters));
+    srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
+                  _mm256_castsi256_si128(firstFilters));
+    srcRegFilt7 = _mm_maddubs_epi16(srcRegFilt7,
+                  _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
+
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+                  _mm256_castsi256_si128(secondFilters));
+    srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
+                  _mm256_castsi256_si128(secondFilters));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
+                  _mm256_castsi256_si128(thirdFilters));
+    srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
+                  _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm_min_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm_min_epi16(srcRegFilt5, srcRegFilt7));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm_max_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm_max_epi16(srcRegFilt5, srcRegFilt7));
+
+
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                  _mm256_castsi256_si128(addFilterReg64));
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3,
+                  _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
+  }
+}
+
+#if HAVE_AVX2 && HAVE_SSSE3
+filter8_1dfunction vp9_filter_block1d4_v8_ssse3;
+#if ARCH_X86_64
+filter8_1dfunction vp9_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_intrin_ssse3;
+#define vp9_filter_block1d8_v8_avx2 vp9_filter_block1d8_v8_intrin_ssse3
+#define vp9_filter_block1d8_h8_avx2 vp9_filter_block1d8_h8_intrin_ssse3
+#define vp9_filter_block1d4_h8_avx2 vp9_filter_block1d4_h8_intrin_ssse3
+#else  // ARCH_X86
+filter8_1dfunction vp9_filter_block1d8_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_ssse3;
+#define vp9_filter_block1d8_v8_avx2 vp9_filter_block1d8_v8_ssse3
+#define vp9_filter_block1d8_h8_avx2 vp9_filter_block1d8_h8_ssse3
+#define vp9_filter_block1d4_h8_avx2 vp9_filter_block1d4_h8_ssse3
+#endif  // ARCH_X86_64
+filter8_1dfunction vp9_filter_block1d16_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h2_ssse3;
+#define vp9_filter_block1d4_v8_avx2 vp9_filter_block1d4_v8_ssse3
+#define vp9_filter_block1d16_v2_avx2 vp9_filter_block1d16_v2_ssse3
+#define vp9_filter_block1d16_h2_avx2 vp9_filter_block1d16_h2_ssse3
+#define vp9_filter_block1d8_v2_avx2  vp9_filter_block1d8_v2_ssse3
+#define vp9_filter_block1d8_h2_avx2  vp9_filter_block1d8_h2_ssse3
+#define vp9_filter_block1d4_v2_avx2  vp9_filter_block1d4_v2_ssse3
+#define vp9_filter_block1d4_h2_avx2  vp9_filter_block1d4_h2_ssse3
+// void vp9_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void vp9_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
+
+// void vp9_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+FUN_CONV_2D(, avx2);
+#endif  // HAVE_AX2 && HAVE_SSSE3
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c
new file mode 100644
index 000000000..5fd2857e1
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c
@@ -0,0 +1,601 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Due to a header conflict between math.h and intrinsics includes with ceil()
+// in certain configurations under vs9 this include needs to precede
+// tmmintrin.h.
+#include "./vp9_rtcd.h"
+
+#include <tmmintrin.h>
+
+#include "vp9/common/x86/convolve.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+// filters only for the 4_h8 convolution
+DECLARE_ALIGNED(16, static const uint8_t, filt1_4_h8[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 2, 3, 3, 4, 4, 5, 5, 6
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_4_h8[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+// filters for 8_h8 and 16_h8
+DECLARE_ALIGNED(16, static const uint8_t, filt1_global[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_global[16]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt3_global[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+// These are reused by the avx2 intrinsics.
+filter8_1dfunction vp9_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_intrin_ssse3;
+
+void vp9_filter_block1d4_h8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i firstFilters, secondFilters, shuffle1, shuffle2;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, srcReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 =_mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter into the first lane
+  firstFilters = _mm_shufflelo_epi16(filtersReg, 0);
+  // duplicate only the third 16 bit in the filter into the first lane
+  secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu);
+  // duplicate only the seconds 16 bits in the filter into the second lane
+  // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3
+  firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u);
+  // duplicate only the forth 16 bits in the filter into the second lane
+  // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7
+  secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu);
+
+  // loading the local filters
+  shuffle1 =_mm_load_si128((__m128i const *)filt1_4_h8);
+  shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, shuffle1);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, shuffle2);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // extract the higher half of the lane
+    srcRegFilt3 =  _mm_srli_si128(srcRegFilt1, 8);
+    srcRegFilt4 =  _mm_srli_si128(srcRegFilt2, 8);
+
+    minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2);
+
+    // add and saturate all the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_max_epi16(srcRegFilt3, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+    src_ptr+=src_pixels_per_line;
+
+    // save only 4 bytes
+    *((int*)&output_ptr[0])= _mm_cvtsi128_si32(srcRegFilt1);
+
+    output_ptr+=output_pitch;
+  }
+}
+
+void vp9_filter_block1d8_h8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pixels_per_line,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t output_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters, srcReg;
+  __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 128 bit register
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 128 bit register
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
+  filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
+  filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
+  filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, filt1Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, filt2Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg, filt3Reg);
+    srcRegFilt4= _mm_shuffle_epi8(srcReg, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters);
+    srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters);
+
+    // add and saturate all the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+
+    srcRegFilt2= _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr+=src_pixels_per_line;
+
+    // save only 8 bytes
+    _mm_storel_epi64((__m128i*)&output_ptr[0], srcRegFilt1);
+
+    output_ptr+=output_pitch;
+  }
+}
+
+static void vp9_filter_block1d16_h8_intrin_ssse3(const uint8_t *src_ptr,
+                                                 ptrdiff_t src_pixels_per_line,
+                                                 uint8_t *output_ptr,
+                                                 ptrdiff_t output_pitch,
+                                                 uint32_t output_height,
+                                                 const int16_t *filter) {
+  __m128i addFilterReg64, filtersReg, srcReg1, srcReg2;
+  __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i srcRegFilt1_1, srcRegFilt2_1, srcRegFilt2, srcRegFilt3;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 128 bit register
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 128 bit register
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
+  filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
+  filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
+  filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1_1= _mm_shuffle_epi8(srcReg1, filt1Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg1, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                                   _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+
+    // reading the next 16 bytes.
+    // (part of it was being read by earlier read)
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
+                                   _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    // filter the source buffer
+    srcRegFilt2_1= _mm_shuffle_epi8(srcReg2, filt1Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3= _mm_shuffle_epi8(srcReg2, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1,
+    _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, addFilterReg64);
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
+    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+    src_ptr+=src_pixels_per_line;
+
+    // save 16 bytes
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1_1);
+
+    output_ptr+=output_pitch;
+  }
+}
+
+void vp9_filter_block1d8_v8_intrin_ssse3(const uint8_t *src_ptr,
+                                         ptrdiff_t src_pitch,
+                                         uint8_t *output_ptr,
+                                         ptrdiff_t out_pitch,
+                                         uint32_t output_height,
+                                         const int16_t *filter) {
+  __m128i addFilterReg64, filtersReg, minReg;
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt5;
+  __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7;
+  __m128i srcReg8;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits in the filter
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits in the filter
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits in the filter
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  // load the first 7 rows of 8 bytes
+  srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr);
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+  srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+
+  for (i = 0; i < output_height; i++) {
+    // load the last 8 bytes
+    srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the result together
+    srcRegFilt1 = _mm_unpacklo_epi8(srcReg1, srcReg2);
+    srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+
+    // merge the result together
+    srcRegFilt2 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcRegFilt5 = _mm_unpacklo_epi8(srcReg7, srcReg8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
+    srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, forthFilters);
+
+    // add and saturate the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt5);
+    srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr+=src_pitch;
+
+    // shift down a row
+    srcReg1 = srcReg2;
+    srcReg2 = srcReg3;
+    srcReg3 = srcReg4;
+    srcReg4 = srcReg5;
+    srcReg5 = srcReg6;
+    srcReg6 = srcReg7;
+    srcReg7 = srcReg8;
+
+    // save only 8 bytes convolve result
+    _mm_storel_epi64((__m128i*)&output_ptr[0], srcRegFilt1);
+
+    output_ptr+=out_pitch;
+  }
+}
+
+static void vp9_filter_block1d16_v8_intrin_ssse3(const uint8_t *src_ptr,
+                                                 ptrdiff_t src_pitch,
+                                                 uint8_t *output_ptr,
+                                                 ptrdiff_t out_pitch,
+                                                 uint32_t output_height,
+                                                 const int16_t *filter) {
+  __m128i addFilterReg64, filtersReg, srcRegFilt1, srcRegFilt3;
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i srcRegFilt5, srcRegFilt6, srcRegFilt7, srcRegFilt8;
+  __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7;
+  __m128i srcReg8;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits in the filter
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits in the filter
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits in the filter
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  // load the first 7 rows of 16 bytes
+  srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr));
+  srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+  srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+  srcReg7 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+
+  for (i = 0; i < output_height; i++) {
+    // load the last 16 bytes
+    srcReg8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the result together
+    srcRegFilt5 = _mm_unpacklo_epi8(srcReg1, srcReg2);
+    srcRegFilt6 = _mm_unpacklo_epi8(srcReg7, srcReg8);
+    srcRegFilt1 = _mm_unpackhi_epi8(srcReg1, srcReg2);
+    srcRegFilt3 = _mm_unpackhi_epi8(srcReg7, srcReg8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, firstFilters);
+    srcRegFilt6 = _mm_maddubs_epi16(srcRegFilt6, forthFilters);
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt5 = _mm_adds_epi16(srcRegFilt5, srcRegFilt6);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3);
+
+    // merge the result together
+    srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+    srcRegFilt6 = _mm_unpackhi_epi8(srcReg3, srcReg4);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt6 = _mm_maddubs_epi16(srcRegFilt6, secondFilters);
+
+    // merge the result together
+    srcRegFilt7 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcRegFilt8 = _mm_unpackhi_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt7 = _mm_maddubs_epi16(srcRegFilt7, thirdFilters);
+    srcRegFilt8 = _mm_maddubs_epi16(srcRegFilt8, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt5 = _mm_adds_epi16(srcRegFilt5,
+                                 _mm_min_epi16(srcRegFilt3, srcRegFilt7));
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                                 _mm_min_epi16(srcRegFilt6, srcRegFilt8));
+
+    // add and saturate the results together
+    srcRegFilt5 = _mm_adds_epi16(srcRegFilt5,
+                                 _mm_max_epi16(srcRegFilt3, srcRegFilt7));
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
+                                 _mm_max_epi16(srcRegFilt6, srcRegFilt8));
+    srcRegFilt5 = _mm_adds_epi16(srcRegFilt5, addFilterReg64);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt5 = _mm_srai_epi16(srcRegFilt5, 7);
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt5, srcRegFilt1);
+
+    src_ptr+=src_pitch;
+
+    // shift down a row
+    srcReg1 = srcReg2;
+    srcReg2 = srcReg3;
+    srcReg3 = srcReg4;
+    srcReg4 = srcReg5;
+    srcReg5 = srcReg6;
+    srcReg6 = srcReg7;
+    srcReg7 = srcReg8;
+
+    // save 16 bytes convolve result
+    _mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
+
+    output_ptr+=out_pitch;
+  }
+}
+
+#if ARCH_X86_64
+filter8_1dfunction vp9_filter_block1d16_v8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_intrin_ssse3;
+#define vp9_filter_block1d16_v8_ssse3 vp9_filter_block1d16_v8_intrin_ssse3
+#define vp9_filter_block1d16_h8_ssse3 vp9_filter_block1d16_h8_intrin_ssse3
+#define vp9_filter_block1d8_v8_ssse3 vp9_filter_block1d8_v8_intrin_ssse3
+#define vp9_filter_block1d8_h8_ssse3 vp9_filter_block1d8_h8_intrin_ssse3
+#define vp9_filter_block1d4_h8_ssse3 vp9_filter_block1d4_h8_intrin_ssse3
+#else  // ARCH_X86
+filter8_1dfunction vp9_filter_block1d16_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h8_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v8_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_ssse3;
+#endif  // ARCH_X86_64
+filter8_1dfunction vp9_filter_block1d16_v8_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h8_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v8_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h8_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v8_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h8_avg_ssse3;
+
+filter8_1dfunction vp9_filter_block1d16_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v2_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h2_ssse3;
+filter8_1dfunction vp9_filter_block1d16_v2_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d16_h2_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d8_v2_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d8_h2_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d4_v2_avg_ssse3;
+filter8_1dfunction vp9_filter_block1d4_h2_avg_ssse3;
+
+// void vp9_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void vp9_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void vp9_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h);
+// void vp9_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, ssse3);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+            ssse3);
+
+// void vp9_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+// void vp9_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+FUN_CONV_2D(, ssse3);
+FUN_CONV_2D(avg_ , ssse3);
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_8t_sse2.asm b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_sse2.asm
new file mode 100644
index 000000000..9dc8d0abb
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_sse2.asm
@@ -0,0 +1,987 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklqdq  xmm0, xmm1
+    punpcklqdq  xmm2, xmm3
+    punpcklqdq  xmm5, xmm4
+    punpcklqdq  xmm6, xmm7
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm2
+    movdqa      k5k4, xmm5
+    movdqa      k6k7, xmm6
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpckldq   xmm6, xmm7
+    punpckldq   xmm2, xmm3
+    punpckldq   xmm5, xmm4
+
+    punpcklbw   xmm0, zero                  ;unpack to word
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+
+    pmullw      xmm0, k0k1                  ;multiply the filter factors
+    pmullw      xmm6, k6k7
+    pmullw      xmm2, k2k3
+    pmullw      xmm5, k5k4
+
+    paddsw      xmm0, xmm6                  ;sum
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm2
+    psrldq      xmm2, 8
+    paddsw      xmm0, xmm5
+    psrldq      xmm5, 8
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+%endm
+
+%macro GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm0
+    punpcklwd   xmm1, xmm1
+    punpcklwd   xmm2, xmm2
+    punpcklwd   xmm3, xmm3
+    punpckhwd   xmm4, xmm4
+    punpckhwd   xmm5, xmm5
+    punpckhwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movdqa      k0,   xmm0                  ;store filter factors on stack
+    movdqa      k1,   xmm1
+    movdqa      k2,   xmm2
+    movdqa      k3,   xmm3
+    movdqa      k4,   xmm4
+    movdqa      k5,   xmm5
+    movdqa      k6,   xmm6
+    movdqa      k7,   xmm7
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro LOAD_VERT_8 1
+    movq        xmm0, [rsi + %1]            ;0
+    movq        xmm1, [rsi + rax + %1]      ;1
+    movq        xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2 + %1]  ;7
+    movq        xmm2, [rsi + rax + %1]      ;2
+    movq        xmm3, [rsi + rax * 2 + %1]  ;3
+    movq        xmm4, [rsi + rdx + %1]      ;4
+    movq        xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro APPLY_FILTER_8 2
+    punpcklbw   xmm0, zero
+    punpcklbw   xmm1, zero
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm7, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+    punpcklbw   xmm3, zero
+    punpcklbw   xmm4, zero
+
+    pmullw      xmm0, k0
+    pmullw      xmm1, k1
+    pmullw      xmm6, k6
+    pmullw      xmm7, k7
+    pmullw      xmm2, k2
+    pmullw      xmm5, k5
+    pmullw      xmm3, k3
+    pmullw      xmm4, k4
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm6
+    paddsw      xmm0, xmm7
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+    paddsw      xmm0, xmm3
+    paddsw      xmm0, xmm4
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi + %2]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi + %2], xmm0
+%endm
+
+;void vp9_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d4_v8_sse2) PRIVATE
+sym(vp9_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d8_v8_sse2) PRIVATE
+sym(vp9_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d16_v8_sse2) PRIVATE
+sym(vp9_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 0, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 1, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d4_h8_sse2) PRIVATE
+sym(vp9_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d8_h8_sse2) PRIVATE
+sym(vp9_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d16_h8_sse2) PRIVATE
+sym(vp9_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(vp9_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm
new file mode 100644
index 000000000..4a5bf1b60
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_8t_ssse3.asm
@@ -0,0 +1,1071 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro VERTx4 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm0, xmm4, 0b              ;k0_k1
+    pshuflw     xmm1, xmm4, 01010101b       ;k2_k3
+    pshuflw     xmm2, xmm4, 10101010b       ;k4_k5
+    pshuflw     xmm3, xmm4, 11111111b       ;k6_k7
+
+    punpcklqdq  xmm0, xmm0
+    punpcklqdq  xmm1, xmm1
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm1
+    pshufd      xmm5, xmm5, 0
+    movdqa      k4k5, xmm2
+    movdqa      k6k7, xmm3
+    movdqa      krd, xmm5
+
+    movsxd      rdx, DWORD PTR arg(1)       ;pixels_per_line
+
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)        ;out_pitch
+%endif
+    mov         rax, rsi
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+    add         rax, rdx
+
+    lea         rbx, [rdx + rdx*4]
+    add         rbx, rdx                    ;pitch * 6
+
+.loop:
+    movd        xmm0, [rsi]                 ;A
+    movd        xmm1, [rsi + rdx]           ;B
+    movd        xmm2, [rsi + rdx * 2]       ;C
+    movd        xmm3, [rax + rdx * 2]       ;D
+    movd        xmm4, [rsi + rdx * 4]       ;E
+    movd        xmm5, [rax + rdx * 4]       ;F
+
+    punpcklbw   xmm0, xmm1                  ;A B
+    punpcklbw   xmm2, xmm3                  ;C D
+    punpcklbw   xmm4, xmm5                  ;E F
+
+    movd        xmm6, [rsi + rbx]           ;G
+    movd        xmm7, [rax + rbx]           ;H
+
+    pmaddubsw   xmm0, k0k1
+    pmaddubsw   xmm2, k2k3
+    punpcklbw   xmm6, xmm7                  ;G H
+    pmaddubsw   xmm4, k4k5
+    pmaddubsw   xmm6, k6k7
+
+    movdqa      xmm1, xmm2
+    paddsw      xmm0, xmm6
+    pmaxsw      xmm2, xmm4
+    pminsw      xmm4, xmm1
+    paddsw      xmm0, xmm4
+    paddsw      xmm0, xmm2
+
+    paddsw      xmm0, krd
+    psraw       xmm0, 7
+    packuswb    xmm0, xmm0
+
+    add         rsi,  rdx
+    add         rax,  rdx
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+
+%if ABI_IS_32BIT
+    add         rdi, DWORD PTR arg(3)       ;out_pitch
+%else
+    add         rdi, r8
+%endif
+    dec         rcx
+    jnz         .loop
+%endm
+
+%macro VERTx8 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm0, xmm4, 0b              ;k0_k1
+    pshuflw     xmm1, xmm4, 01010101b       ;k2_k3
+    pshuflw     xmm2, xmm4, 10101010b       ;k4_k5
+    pshuflw     xmm3, xmm4, 11111111b       ;k6_k7
+
+    punpcklqdq  xmm0, xmm0
+    punpcklqdq  xmm1, xmm1
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm1
+    pshufd      xmm5, xmm5, 0
+    movdqa      k4k5, xmm2
+    movdqa      k6k7, xmm3
+    movdqa      krd, xmm5
+
+    movsxd      rdx, DWORD PTR arg(1)       ;pixels_per_line
+
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)        ;out_pitch
+%endif
+    mov         rax, rsi
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+    add         rax, rdx
+
+    lea         rbx, [rdx + rdx*4]
+    add         rbx, rdx                    ;pitch * 6
+
+.loop:
+    movq        xmm0, [rsi]                 ;A
+    movq        xmm1, [rsi + rdx]           ;B
+    movq        xmm2, [rsi + rdx * 2]       ;C
+    movq        xmm3, [rax + rdx * 2]       ;D
+    movq        xmm4, [rsi + rdx * 4]       ;E
+    movq        xmm5, [rax + rdx * 4]       ;F
+
+    punpcklbw   xmm0, xmm1                  ;A B
+    punpcklbw   xmm2, xmm3                  ;C D
+    punpcklbw   xmm4, xmm5                  ;E F
+
+    movq        xmm6, [rsi + rbx]           ;G
+    movq        xmm7, [rax + rbx]           ;H
+
+    pmaddubsw   xmm0, k0k1
+    pmaddubsw   xmm2, k2k3
+    punpcklbw   xmm6, xmm7                  ;G H
+    pmaddubsw   xmm4, k4k5
+    pmaddubsw   xmm6, k6k7
+
+    paddsw      xmm0, xmm6
+    movdqa      xmm1, xmm2
+    pmaxsw      xmm2, xmm4
+    pminsw      xmm4, xmm1
+    paddsw      xmm0, xmm4
+    paddsw      xmm0, xmm2
+
+    paddsw      xmm0, krd
+    psraw       xmm0, 7
+    packuswb    xmm0, xmm0
+
+    add         rsi,  rdx
+    add         rax,  rdx
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+
+%if ABI_IS_32BIT
+    add         rdi, DWORD PTR arg(3)       ;out_pitch
+%else
+    add         rdi, r8
+%endif
+    dec         rcx
+    jnz         .loop
+%endm
+
+
+%macro VERTx16 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm0, xmm4, 0b              ;k0_k1
+    pshuflw     xmm1, xmm4, 01010101b       ;k2_k3
+    pshuflw     xmm2, xmm4, 10101010b       ;k4_k5
+    pshuflw     xmm3, xmm4, 11111111b       ;k6_k7
+
+    punpcklqdq  xmm0, xmm0
+    punpcklqdq  xmm1, xmm1
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm1
+    pshufd      xmm5, xmm5, 0
+    movdqa      k4k5, xmm2
+    movdqa      k6k7, xmm3
+    movdqa      krd, xmm5
+
+    movsxd      rdx, DWORD PTR arg(1)       ;pixels_per_line
+
+%if ABI_IS_32BIT=0
+    movsxd      r8, DWORD PTR arg(3)        ;out_pitch
+%endif
+    mov         rax, rsi
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+    add         rax, rdx
+
+    lea         rbx, [rdx + rdx*4]
+    add         rbx, rdx                    ;pitch * 6
+
+.loop:
+    movq        xmm0, [rsi]                 ;A
+    movq        xmm1, [rsi + rdx]           ;B
+    movq        xmm2, [rsi + rdx * 2]       ;C
+    movq        xmm3, [rax + rdx * 2]       ;D
+    movq        xmm4, [rsi + rdx * 4]       ;E
+    movq        xmm5, [rax + rdx * 4]       ;F
+
+    punpcklbw   xmm0, xmm1                  ;A B
+    punpcklbw   xmm2, xmm3                  ;C D
+    punpcklbw   xmm4, xmm5                  ;E F
+
+    movq        xmm6, [rsi + rbx]           ;G
+    movq        xmm7, [rax + rbx]           ;H
+
+    pmaddubsw   xmm0, k0k1
+    pmaddubsw   xmm2, k2k3
+    punpcklbw   xmm6, xmm7                  ;G H
+    pmaddubsw   xmm4, k4k5
+    pmaddubsw   xmm6, k6k7
+
+    paddsw      xmm0, xmm6
+    movdqa      xmm1, xmm2
+    pmaxsw      xmm2, xmm4
+    pminsw      xmm4, xmm1
+    paddsw      xmm0, xmm4
+    paddsw      xmm0, xmm2
+
+    paddsw      xmm0, krd
+    psraw       xmm0, 7
+    packuswb    xmm0, xmm0
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+
+    movq        xmm0, [rsi + 8]             ;A
+    movq        xmm1, [rsi + rdx + 8]       ;B
+    movq        xmm2, [rsi + rdx * 2 + 8]   ;C
+    movq        xmm3, [rax + rdx * 2 + 8]   ;D
+    movq        xmm4, [rsi + rdx * 4 + 8]   ;E
+    movq        xmm5, [rax + rdx * 4 + 8]   ;F
+
+    punpcklbw   xmm0, xmm1                  ;A B
+    punpcklbw   xmm2, xmm3                  ;C D
+    punpcklbw   xmm4, xmm5                  ;E F
+
+    movq        xmm6, [rsi + rbx + 8]       ;G
+    movq        xmm7, [rax + rbx + 8]       ;H
+    punpcklbw   xmm6, xmm7                  ;G H
+
+    pmaddubsw   xmm0, k0k1
+    pmaddubsw   xmm2, k2k3
+    pmaddubsw   xmm4, k4k5
+    pmaddubsw   xmm6, k6k7
+
+    paddsw      xmm0, xmm6
+    movdqa      xmm1, xmm2
+    pmaxsw      xmm2, xmm4
+    pminsw      xmm4, xmm1
+    paddsw      xmm0, xmm4
+    paddsw      xmm0, xmm2
+
+    paddsw      xmm0, krd
+    psraw       xmm0, 7
+    packuswb    xmm0, xmm0
+
+    add         rsi,  rdx
+    add         rax,  rdx
+%if %1
+    movq    xmm1, [rdi+8]
+    pavgb   xmm0, xmm1
+%endif
+
+    movq        [rdi+8], xmm0
+
+%if ABI_IS_32BIT
+    add         rdi, DWORD PTR arg(3)       ;out_pitch
+%else
+    add         rdi, r8
+%endif
+    dec         rcx
+    jnz         .loop
+%endm
+
+;void vp9_filter_block1d8_v8_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d4_v8_ssse3) PRIVATE
+sym(vp9_filter_block1d4_v8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx4 0
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_v8_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d8_v8_ssse3) PRIVATE
+sym(vp9_filter_block1d8_v8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx8 0
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_v8_ssse3
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d16_v8_ssse3) PRIVATE
+sym(vp9_filter_block1d16_v8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx16 0
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+
+global sym(vp9_filter_block1d4_v8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d4_v8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx4 1
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d8_v8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx8 1
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d16_v8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    VERTx16 1
+
+    add rsp, 16*5
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+%macro HORIZx4_ROW 2
+    movdqa      %2,   %1
+    pshufb      %1,   [GLOBAL(shuf_t0t1)]
+    pshufb      %2,   [GLOBAL(shuf_t2t3)]
+    pmaddubsw   %1,   k0k1k4k5
+    pmaddubsw   %2,   k2k3k6k7
+
+    movdqa      xmm4, %1
+    movdqa      xmm5, %2
+    psrldq      %1,   8
+    psrldq      %2,   8
+    movdqa      xmm6, xmm5
+
+    paddsw      xmm4, %2
+    pmaxsw      xmm5, %1
+    pminsw      %1, xmm6
+    paddsw      %1, xmm4
+    paddsw      %1, xmm5
+
+    paddsw      %1,   krd
+    psraw       %1,   7
+    packuswb    %1,   %1
+%endm
+
+%macro HORIZx4 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm6, xmm4, 0b              ;k0_k1
+    pshufhw     xmm6, xmm6, 10101010b       ;k0_k1_k4_k5
+    pshuflw     xmm7, xmm4, 01010101b       ;k2_k3
+    pshufhw     xmm7, xmm7, 11111111b       ;k2_k3_k6_k7
+    pshufd      xmm5, xmm5, 0               ;rounding
+
+    movdqa      k0k1k4k5, xmm6
+    movdqa      k2k3k6k7, xmm7
+    movdqa      krd, xmm5
+
+    movsxd      rax, dword ptr arg(1)       ;src_pixels_per_line
+    movsxd      rdx, dword ptr arg(3)       ;output_pitch
+    movsxd      rcx, dword ptr arg(4)       ;output_height
+    shr         rcx, 1
+.loop:
+    ;Do two rows once
+    movq        xmm0,   [rsi - 3]           ;load src
+    movq        xmm1,   [rsi + 5]
+    movq        xmm2,   [rsi + rax - 3]
+    movq        xmm3,   [rsi + rax + 5]
+    punpcklqdq  xmm0,   xmm1
+    punpcklqdq  xmm2,   xmm3
+
+    HORIZx4_ROW xmm0,   xmm1
+    HORIZx4_ROW xmm2,   xmm3
+%if %1
+    movd        xmm1,   [rdi]
+    pavgb       xmm0,   xmm1
+    movd        xmm3,   [rdi + rdx]
+    pavgb       xmm2,   xmm3
+%endif
+    movd        [rdi],  xmm0
+    movd        [rdi +rdx],  xmm2
+
+    lea         rsi,    [rsi + rax]
+    prefetcht0  [rsi + 4 * rax - 3]
+    lea         rsi,    [rsi + rax]
+    lea         rdi,    [rdi + 2 * rdx]
+    prefetcht0  [rsi + 2 * rax - 3]
+
+    dec         rcx
+    jnz         .loop
+
+    ; Do last row if output_height is odd
+    movsxd      rcx,    dword ptr arg(4)       ;output_height
+    and         rcx,    1
+    je          .done
+
+    movq        xmm0,   [rsi - 3]    ; load src
+    movq        xmm1,   [rsi + 5]
+    punpcklqdq  xmm0,   xmm1
+
+    HORIZx4_ROW xmm0, xmm1
+%if %1
+    movd        xmm1,   [rdi]
+    pavgb       xmm0,   xmm1
+%endif
+    movd        [rdi],  xmm0
+.done
+%endm
+
+%macro HORIZx8_ROW 4
+    movdqa      %2,   %1
+    movdqa      %3,   %1
+    movdqa      %4,   %1
+
+    pshufb      %1,   [GLOBAL(shuf_t0t1)]
+    pshufb      %2,   [GLOBAL(shuf_t2t3)]
+    pshufb      %3,   [GLOBAL(shuf_t4t5)]
+    pshufb      %4,   [GLOBAL(shuf_t6t7)]
+
+    pmaddubsw   %1,   k0k1
+    pmaddubsw   %2,   k2k3
+    pmaddubsw   %3,   k4k5
+    pmaddubsw   %4,   k6k7
+
+    paddsw      %1,   %4
+    movdqa      %4,   %2
+    pmaxsw      %2,   %3
+    pminsw      %3,   %4
+    paddsw      %1,   %3
+    paddsw      %1,   %2
+
+    paddsw      %1,   krd
+    psraw       %1,   7
+    packuswb    %1,   %1
+%endm
+
+%macro HORIZx8 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm0, xmm4, 0b              ;k0_k1
+    pshuflw     xmm1, xmm4, 01010101b       ;k2_k3
+    pshuflw     xmm2, xmm4, 10101010b       ;k4_k5
+    pshuflw     xmm3, xmm4, 11111111b       ;k6_k7
+
+    punpcklqdq  xmm0, xmm0
+    punpcklqdq  xmm1, xmm1
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm1
+    pshufd      xmm5, xmm5, 0
+    movdqa      k4k5, xmm2
+    movdqa      k6k7, xmm3
+    movdqa      krd, xmm5
+
+    movsxd      rax, dword ptr arg(1)       ;src_pixels_per_line
+    movsxd      rdx, dword ptr arg(3)       ;output_pitch
+    movsxd      rcx, dword ptr arg(4)       ;output_height
+    shr         rcx, 1
+
+.loop:
+    movq        xmm0,   [rsi - 3]           ;load src
+    movq        xmm3,   [rsi + 5]
+    movq        xmm4,   [rsi + rax - 3]
+    movq        xmm7,   [rsi + rax + 5]
+    punpcklqdq  xmm0,   xmm3
+    punpcklqdq  xmm4,   xmm7
+
+    HORIZx8_ROW xmm0, xmm1, xmm2, xmm3
+    HORIZx8_ROW xmm4, xmm5, xmm6, xmm7
+%if %1
+    movq        xmm1,   [rdi]
+    movq        xmm2,   [rdi + rdx]
+    pavgb       xmm0,   xmm1
+    pavgb       xmm4,   xmm2
+%endif
+    movq        [rdi],  xmm0
+    movq        [rdi + rdx],  xmm4
+
+    lea         rsi,    [rsi + rax]
+    prefetcht0  [rsi + 4 * rax - 3]
+    lea         rsi,    [rsi + rax]
+    lea         rdi,    [rdi + 2 * rdx]
+    prefetcht0  [rsi + 2 * rax - 3]
+    dec         rcx
+    jnz         .loop
+
+    ;Do last row if output_height is odd
+    movsxd      rcx,    dword ptr arg(4)    ;output_height
+    and         rcx,    1
+    je          .done
+
+    movq        xmm0,   [rsi - 3]
+    movq        xmm3,   [rsi + 5]
+    punpcklqdq  xmm0,   xmm3
+
+    HORIZx8_ROW xmm0, xmm1, xmm2, xmm3
+%if %1
+    movq        xmm1,   [rdi]
+    pavgb       xmm0,   xmm1
+%endif
+    movq        [rdi],  xmm0
+.done
+%endm
+
+%macro HORIZx16 1
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm4, [rdx]                 ;load filters
+    movq        xmm5, rcx
+    packsswb    xmm4, xmm4
+    pshuflw     xmm0, xmm4, 0b              ;k0_k1
+    pshuflw     xmm1, xmm4, 01010101b       ;k2_k3
+    pshuflw     xmm2, xmm4, 10101010b       ;k4_k5
+    pshuflw     xmm3, xmm4, 11111111b       ;k6_k7
+
+    punpcklqdq  xmm0, xmm0
+    punpcklqdq  xmm1, xmm1
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm1
+    pshufd      xmm5, xmm5, 0
+    movdqa      k4k5, xmm2
+    movdqa      k6k7, xmm3
+    movdqa      krd, xmm5
+
+    movsxd      rax, dword ptr arg(1)       ;src_pixels_per_line
+    movsxd      rdx, dword ptr arg(3)       ;output_pitch
+    movsxd      rcx, dword ptr arg(4)       ;output_height
+
+.loop:
+    prefetcht0  [rsi + 2 * rax -3]
+
+    movq        xmm0,   [rsi - 3]           ;load src data
+    movq        xmm4,   [rsi + 5]
+    movq        xmm6,   [rsi + 13]
+    punpcklqdq  xmm0,   xmm4
+    punpcklqdq  xmm4,   xmm6
+
+    movdqa      xmm7,   xmm0
+
+    punpcklbw   xmm7,   xmm7
+    punpckhbw   xmm0,   xmm0
+    movdqa      xmm1,   xmm0
+    movdqa      xmm2,   xmm0
+    movdqa      xmm3,   xmm0
+
+    palignr     xmm0,   xmm7, 1
+    palignr     xmm1,   xmm7, 5
+    pmaddubsw   xmm0,   k0k1
+    palignr     xmm2,   xmm7, 9
+    pmaddubsw   xmm1,   k2k3
+    palignr     xmm3,   xmm7, 13
+
+    pmaddubsw   xmm2,   k4k5
+    pmaddubsw   xmm3,   k6k7
+    paddsw      xmm0,   xmm3
+
+    movdqa      xmm3,   xmm4
+    punpcklbw   xmm3,   xmm3
+    punpckhbw   xmm4,   xmm4
+
+    movdqa      xmm5,   xmm4
+    movdqa      xmm6,   xmm4
+    movdqa      xmm7,   xmm4
+
+    palignr     xmm4,   xmm3, 1
+    palignr     xmm5,   xmm3, 5
+    palignr     xmm6,   xmm3, 9
+    palignr     xmm7,   xmm3, 13
+
+    movdqa      xmm3,   xmm1
+    pmaddubsw   xmm4,   k0k1
+    pmaxsw      xmm1,   xmm2
+    pmaddubsw   xmm5,   k2k3
+    pminsw      xmm2,   xmm3
+    pmaddubsw   xmm6,   k4k5
+    paddsw      xmm0,   xmm2
+    pmaddubsw   xmm7,   k6k7
+    paddsw      xmm0,   xmm1
+
+    paddsw      xmm4,   xmm7
+    movdqa      xmm7,   xmm5
+    pmaxsw      xmm5,   xmm6
+    pminsw      xmm6,   xmm7
+    paddsw      xmm4,   xmm6
+    paddsw      xmm4,   xmm5
+
+    paddsw      xmm0,   krd
+    paddsw      xmm4,   krd
+    psraw       xmm0,   7
+    psraw       xmm4,   7
+    packuswb    xmm0,   xmm0
+    packuswb    xmm4,   xmm4
+    punpcklqdq  xmm0,   xmm4
+%if %1
+    movdqa      xmm1,   [rdi]
+    pavgb       xmm0,   xmm1
+%endif
+
+    lea         rsi,    [rsi + rax]
+    movdqa      [rdi],  xmm0
+
+    lea         rdi,    [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+%endm
+
+;void vp9_filter_block1d4_h8_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d4_h8_ssse3) PRIVATE
+sym(vp9_filter_block1d4_h8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 3
+    %define k0k1k4k5 [rsp + 16 * 0]
+    %define k2k3k6k7 [rsp + 16 * 1]
+    %define krd      [rsp + 16 * 2]
+
+    HORIZx4 0
+
+    add rsp, 16 * 3
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d8_h8_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d8_h8_ssse3) PRIVATE
+sym(vp9_filter_block1d8_h8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    HORIZx8 0
+
+    add rsp, 16*5
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void vp9_filter_block1d16_h8_ssse3
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(vp9_filter_block1d16_h8_ssse3) PRIVATE
+sym(vp9_filter_block1d16_h8_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    HORIZx16 0
+
+    add rsp, 16*5
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d4_h8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 3
+    %define k0k1k4k5 [rsp + 16 * 0]
+    %define k2k3k6k7 [rsp + 16 * 1]
+    %define krd      [rsp + 16 * 2]
+
+    HORIZx4 1
+
+    add rsp, 16 * 3
+    pop rsp
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d8_h8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    HORIZx8 1
+
+    add rsp, 16*5
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h8_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d16_h8_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16*5
+    %define k0k1 [rsp + 16*0]
+    %define k2k3 [rsp + 16*1]
+    %define k4k5 [rsp + 16*2]
+    %define k6k7 [rsp + 16*3]
+    %define krd [rsp + 16*4]
+
+    HORIZx16 1
+
+    add rsp, 16*5
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+SECTION_RODATA
+align 16
+shuf_t0t1:
+    db  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+align 16
+shuf_t2t3:
+    db  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+align 16
+shuf_t4t5:
+    db  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+align 16
+shuf_t6t7:
+    db  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_sse2.asm b/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_sse2.asm
new file mode 100644
index 000000000..d94ccf2e9
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_sse2.asm
@@ -0,0 +1,448 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklqdq  xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    pxor        xmm2, xmm2
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpcklbw   xmm0, xmm2                  ;unpack to word
+    pmullw      xmm0, xmm4                  ;multiply the filter factors
+
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+
+    paddsw      xmm0, xmm3                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+
+    pshuflw     xmm6, xmm7, 11111111b       ;k3
+    pshufhw     xmm7, xmm7, 0b              ;k4
+    punpcklwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    pxor        xmm5, xmm5
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm4                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+    punpckhbw   xmm2, xmm5
+    punpckhbw   xmm3, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    pmullw      xmm2, xmm6
+    pmullw      xmm3, xmm7
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm2, xmm3
+
+    paddsw      xmm0, xmm4                  ;rounding
+    paddsw      xmm2, xmm4
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(vp9_filter_block1d4_v2_sse2) PRIVATE
+sym(vp9_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v2_sse2) PRIVATE
+sym(vp9_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v2_sse2) PRIVATE
+sym(vp9_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h2_sse2) PRIVATE
+sym(vp9_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h2_sse2) PRIVATE
+sym(vp9_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h2_sse2) PRIVATE
+sym(vp9_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(vp9_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_ssse3.asm b/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_ssse3.asm
new file mode 100644
index 000000000..b5e18fe6d
--- /dev/null
+++ b/media/libvpx/vp9/common/x86/vp9_subpixel_bilinear_ssse3.asm
@@ -0,0 +1,422 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    psrldq      xmm3, 6
+    packsswb    xmm3, xmm3
+    pshuflw     xmm3, xmm3, 0b              ;k3_k4
+
+    movq        xmm2, rcx                   ;rounding
+    pshufd      xmm2, xmm2, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm3
+
+    paddsw      xmm0, xmm2                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    psrldq      xmm7, 6
+    packsswb    xmm7, xmm7
+    pshuflw     xmm7, xmm7, 0b              ;k3_k4
+    punpcklwd   xmm7, xmm7
+
+    movq        xmm6, rcx                   ;rounding
+    pshufd      xmm6, xmm6, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm7
+
+    paddsw      xmm0, xmm6                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm1
+    punpckhbw   xmm2, xmm1
+    pmaddubsw   xmm0, xmm7
+    pmaddubsw   xmm2, xmm7
+
+    paddsw      xmm0, xmm6                  ;rounding
+    paddsw      xmm2, xmm6
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(vp9_filter_block1d4_v2_ssse3) PRIVATE
+sym(vp9_filter_block1d4_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v2_ssse3) PRIVATE
+sym(vp9_filter_block1d8_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v2_ssse3) PRIVATE
+sym(vp9_filter_block1d16_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_v2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d4_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_v2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d8_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_v2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d16_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h2_ssse3) PRIVATE
+sym(vp9_filter_block1d4_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h2_ssse3) PRIVATE
+sym(vp9_filter_block1d8_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h2_ssse3) PRIVATE
+sym(vp9_filter_block1d16_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d4_h2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d4_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d8_h2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d8_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(vp9_filter_block1d16_h2_avg_ssse3) PRIVATE
+sym(vp9_filter_block1d16_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret